《Artefact:生成式人工智能于医疗健康(2023)(英文版)(23页).pdf》由会员分享,可在线阅读,更多相关《Artefact:生成式人工智能于医疗健康(2023)(英文版)(23页).pdf(23页珍藏版)》请在三个皮匠报告上搜索。
1、Unlocking the potential of Generative AI for patients,practitioners and pharmaceutical companies.IN PARTNERSHIP WITHGENERATIVE AI REPORT FOR HEALTHCARE3GENERATIVE AI FOR HEALTHCARE12345TABLE OF CONTENTSWE UNLEASH THE FULL VALUE OF YOUR PEOPLE AND BUSINESSES THROUGH DATA AND AI SOLUTIONS.18COUNTRIES+
2、1000CLIENTS+1300EMPLOYEESDATA ACCELERATION PROGRAMS|AI SOLUTIONS|DATA MARKETINGArtefact is a global leader in consulting services,specialized in data transformation and data&digital marketing,from strategy to the deployment of AI solutions.We offer a unique combination of innovation(Art)and data sci
3、ence(Fact).The rise of Generative AI in healthcare:Navigating between promise and controlConcrete Generative AI healthcare:Applications and benefitsThe healthcare ecosystem:Gearing upto unlock the potential of Generative AIGenerative AI in healthcare:Limitations,challenges and opportunitiesTrust and
4、 control:A critical role in realizing the potential of Generative AI in healthcareUnlocking the potential of Generative AI for patients,practitioners and pharmaceutical companies.GENERATIVE AI REPORT FOR HEALTHCARE040718334045GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCAREGenerative AI has
5、experienced massive adoption and has gained significant momentum in 2023.While its prospects for healthcare are promising,the biggest challenge will be our ability to control outcomes.We are witnessing unprecedented hypergrowth in medical knowledge:while it doubled every 50 years in the 1950s,it acc
6、elerated to 3.5 years in 2010 and then to every 73 days in 2020.Knowledge is expanding faster than our ability to consume it,both in patient care and research.At the same time,the patient population is growing and complex pathologies are becoming more common,leading healthcare professionals toward h
7、yperspecialization.These phenomena put significant pressure on healthcare professionals,who are overwhelmed by the content while having less time to assimilate it.Therefore,it is crucial to leverage technology to support them.Generative AI holds the transformative potential to liberate humans from r
8、epetitive tasks,allowing them to focus their efforts on higher-value activities and freeing up time to address more complex needs.It can automatically summarize data regardless of volume.It facilitates output customization,allowing to refine results,modify their form,language.It supports the creativ
9、e process by quickly generating ideas and different alternatives of the same result.The impact on the industry is estimated at$1 trillion(McKinsey:Tackling Healthcares Biggest Burdens with Generative AI).We foresee impacts across the entire healthcare value chain,which are developed in the first sec
10、tion of this white paper.However,these opportunities come with significant challenges.If we set aside the traditional issues we face with AI(ethical,regulatory.),two challenges remain.The first is the potential for Generative AI models to hallucinate.Most large language models are based on a specifi
11、c neural network technology called a transformer.This model is trained to identify patterns,to make connections between concepts,to predict words one after another.This partly explains hallucinations:nonsensical sentences of completely fabricated information.This The rise of Generative AI in Healthc
12、are:Navigating between promise and controlFOREWORDis a major concern for some applications that directly affect the health of patients(e.g.,diagnostic support,patient information.).Therefore,a key success factor for Generative AI applications in healthcare will be the validation and control processe
13、s.The second is complementary to the first:these hallucinations appear so plausible that it becomes very challenging to distinguish between whats true and whats false.For example,Chat GPT results often appear to be perfect in form,well written and presented,with arguments that seem almost irrefutabl
14、e-but are not necessarily true.Results like these naturally inspire trust for end users who have not been forewarned.Therefore,the second key success factor for the application of Generative AI will be the training of healthcare professionals and patients.Every end user should know what to expect fr
15、om these applications and what their limitations are.Earlier this year,Artefact released a Generative AI survey describing the technology,the rewards and the risks.This comprehensive survey covers the applications,technical challenges,human impacts and ethical stakes across industries.Given the spec
16、ificities of the healthcare industry,we decided to write this white paper in order to answer the following questions:How can Generative AI transform the healthcare industry,and what will be its first applications?Who are the main actors contributing to the spread of Generative AI,and what are their
17、current areas of focus?What are the key challenges in implementing Generative AI applications and what current solutions are being explored?Generative AI represents a profound paradigm shift in the way we work today,as models switch from being task-focused to outcome-driven.Thanks to foundational mo
18、dels,machines are now making an effort to understand humans and transform the way they consume data.Here is a short list of interesting prompts that healthcare professionals might consider using to enhance their daily activities:1What gene mutations should be monitored to diagnose a cutaneous melano
19、ma?Can you find the AEM labels for adverse events related to the use of carbamazepine?List all relevant medical history of the patient to consider for the appendectomy operation.What are the preparatory steps to be taken before my inguinal hernia surgery?Paul de BalincourtDirector-Healthcare Data&AI
20、 Transformation1 FOREWORD1 FOREWORD67GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCAREThe aim of this work is not to be exhaustive,but rather to highlight disruptive Generative AI use cases and players within the healthcare ecosystem that have captured our interest.Our methodologyTo write thi
21、s white paper,we took an ecosystem-focused approach and categorized the Generative AI healthcare landscape into two main groups:patient-facing actors and enablers.We chose to place the patient at the center of our analysis because we believe that the interests and safety of patients should always co
22、me first when considering the launch of a Generative AI initiative in the healthcare and pharmaceutical industries.For each type of actor,we look at current and future use cases for Generative AI,how they are gearing up to develop these applications,and the limitations,challenges,and opportunities t
23、hey may face.Concrete Generative AI healthcare:Applications and benefits21 FOREWORD89GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCAREGenerative AI is already impacting the entire healthcare ecosystemApplied to the healthcare industry,Generative AI has the potential to deliver significant sho
24、rt-term improvements in operations management,especially through the development of virtual assistants and human augmentation for researchers and caregivers.In the medium to long term,it should drive real innovation acceleration in high value-added areas such as drug discovery,precision medicine,and
25、 care decision making.The rapid and widespread adoption of Generative AI solutions in 2023 has made it challenging to synthesize all potential uses.Nonetheless,several prominent categories of use cases seem to emerge:Data augmentation:Generate new data to enrich/expand the datasets used to train AI-
26、based medical devices or validate clinical trials.For some advanced applications of Generative AI,it is critical to keep scientific expertise at the core of the process.For example,we know that AI can be a great ally in the drug discovery process for target identification and compound design.However
27、,in the preclinical development phase,pharmacology and toxicology studies are essential to validate the identified leads and optimize them for efficacy,safety and better drug-like properties.Biological and chemical knowledge remain at the core of development.It is difficult to predict where and when
28、 these use cases will impact the industry.Without making predictions,we can estimate that the following factors will determine the time-to-market of these Generative AI-enabled solutions:Technological maturity:The current level of readiness of technology assets(e.g.,foundational models available and
29、 performant for a dedicated application).Adoption effort:Depending on the issue we are trying to address,the challenges of reliability,validation,and user acceptance vary(e.g.,applications delivered directly to patients face significant adoption challenges).Data readiness:The ability to capture and
30、leverage enough qualitative data to feed the model.It is hard to believe that Generative AI will be able to create successful new therapeutic programs without a deep understanding of biology.Thomas Clozel-CEO and Co-founder-OWKIN Insight generation:Summarize,extract key information from data to supp
31、ort diagnostics and drug discovery.Biomolecule generation:Design new drug candidates from scratch by generating and optimizing de novo biomolecular structures with the desired properties for drug development.Content personalization:Analyze interactions and create personalized content/experiences to
32、enhance communication with patients and HCPs.Productivity and automation:Enhance day-to-day operational tasks such as assisting with code development or completing administrative tasks(e.g.,prescription forms,office visit reports,social security forms).2 APPLICATIONS&BENEFITS2 APPLICATIONS&BENEFITS1
33、011GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCAREPHARMACEUTICAL COMPANIES1-Synthetic patient data generation for clinical trials:Use Generative AI models to reduce the number of patients to recruit(e.g.,for rare diseases or for risk populations).2-De novo biomolecule generation:Craft new p
34、rotein structures thanks to diffusion models to accelerate drug development and improve gene therapy.3-HCP engagement assistant for sales reps:Generate optimal responses for sales reps based on phone calls and emails to better engage HCPs.4-Marketing content generation:Use Generative AI applications
35、 to create different adaptations(e.g.,messages,formats)of a specific marketing content.5-Regulatory conversational assistant:Build a tool to enable regulatory teams to interactively access regulatory resources by country and receive reminders and checklists at each step of the drug approval process.
36、RESEARCH INSTITUTES11-Writing and formatting scientific publications:Accelerate literature review,identify research questions,provide an overview of the current state of the field,and automate formatting and language review.12-Patient pre-screening:Scan a multitude of health and medical records to s
37、treamline the recruitment funnel and better identify appropriate candidates for clinical trials.PAYERS15-Preventive and informational agent for patients:build informational conversational agents to provide information on medical conditions,surgeries,medical procedures,medications,preventive care and
38、 share patient education material.16-Claims processing streamlining:Automate some routine tasks,such as data entry and analysis.Summarize and prioritize claims.STARTUPSStartups can address all the mentioned use cases.PUBLIC HEALTH AGENCIES AND GOVERNMENT13-Public health and resource monitoring:Detec
39、t early signs of outbreaks,monitor and predict the spread of pathogens,and identify sources of infection by analyzing vast amounts of population data.14-Mental health support:Improve mood monitoring,detect changes in human behavior through text,images or voice,and characterize early symptoms of depr
40、ession or anxiety.CARE PROVIDERS 6-Biomarker analysis support:Automate biomarker data processing(text,images)with Generative AI to enable remote patient monitoring and thereby reduce the number of invasive examinations(e.g.,biopsies).7-Image enhancement and analysis for diagnosis and treatment plann
41、ing:Enhance medical imaging by generating synthetic images,improving reconstruction and segmentation and facilitating disease diagnosis and treatment planning.8-Diagnosis and care decision-making support:Generate evidence-based,explainable next steps in a diagnosis or car decision-making context wit
42、h relevant sources available for review.9-HCP administrative assistant:Manage the summarizing of new medical content(e.g.,studies)and writing for administrative tasks(e.g.,taking minutes of patient appointments,writing prescriptions,or updating patient records).10-Medical coding:Provide real-time co
43、ding suggestions and recommendations,ensuring accuracy and compliance by offering insights into appropriate codes,modifiers and documentation based on industry standards.In the example below,we propose a holistic view of the main Generative AI use cases that can be expected in the healthcare and pha
44、rmaceutical domains:2 APPLICATIONS&BENEFITS2 APPLICATIONS&BENEFITS1213GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCAREGenerative AI can reduce the time needed for the third phases of clinical trials,thanks to augmented cohorts(i.e.,virtual patients generated by AI)even though physician valid
45、ation is required at every stage of the process.Stphanie Allassonnire Professor and Vice-President Valorisation and industrial partnerships UNIVERSIT PARIS CITUSE CASE#1TARGETClinical Trial Investigator:As a clinical trial investigator,I oversee and conduct clinical trials to evaluate new medical tr
46、eatments and interventions,ensuring they meet ethical and regulatory standards.CONTEXTOne major hurdle in clinical trials1 is patient recruitment.As the phase III of clinical trials usually requires a large volume of patients(thousands of patients are sometimes needed),it can take up to four years t
47、o gather a sufficiently large group2 to launch this phase.In addition,even once recruitment is completed,it does not ensure the retention of patients3 afterwards.CHALLENGE Accelerate patient recruitment to reduce the time to phase III launch.Facilitate the recruitment of multicentric cohort recruitm
48、ent for complex populations such as rare diseases and high-risk patients.Reduce potential retention difficulties.SOLUTION Centralize all past and current patient journey data4 from your current cohort.Implement a generative adversarial network model5 on your cohort data set to virtually augment it w
49、ith synthetic patient data,including information on clinical features,genomics,treatment and outcomes.Leverage a validation framework to assess fidelity and privacy preservability of the generated data.RESULTS Solve real patient data imbalance and incompleteness issues.Preserve patient privacy as th
50、e generated data are neither a copy nor a representation of the real data.Accelerate the clinical trials process and reduce the cost of patient recruitment.Synthetic patient data generation to accelerate clinical trialsTransforming healthcare:Zoom in on several high-impact use cases of Generative AI
51、In the following section,we present a selection of flagship use cases that are already available or in development and are likely to have a near-to mid-term impact on the pharmaceutical and healthcare industries.NOTEPlease be advised that all content created with Generative AI is subject to a mandat
52、ory human-in-the-loop approach and rigorous assessment for potential risks and compliance with applicable standards and regulations.2 APPLICATIONS&BENEFITS1415GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCAREGenerative AI has the advantage of embracing the complexity of a situation;but in the
53、 end,its up to the human to be able to make the decision;this prevents the patient from feeling perceived only as a data set.Eric VibertProfessor and MD,PhD,IKO,Liver Surgeon in Paul Brousse Hospital AP-HP,Paris Saclay University CHAIRMAN OF BOPA USE CASE#2RESULTS Streamline and improve global quali
54、ty of care by helping physicians more systematically implement regulatory recommendations and standards of care.Enable medical practitioners to spend more time with patients.Personalized care recommendation supportIn 2020,over 100,000 articles were published on a single pathology:COVID.Generative AI
55、 has the potential to relieve healthcare professionals who lack the time to keep up with the ever-expanding volume of scientific literature by providing them with generated summaries of publications.Grgoire Pign CEO-Oncologist and Radiation TherapistPULSELIFEUSE CASE#3TARGETHCP:As an HCP,I am a memb
56、er of the medical,dental,pharmacy or nursing team.I interact with patients and I may prescribe,purchase,supply,recommend or administer medical treatments and products.CONTEXTAs a healthcare practitioner,staying up to date with medical research is a challenge.The exponential growth of medical knowled
57、ge makes it difficult for healthcare professionals to keep up with every new piece of information.This struggle can have a direct impact on the quality of care provided to patients.CHALLENGE Find useful and relevant content among numerous sources(studies,clinical guidelines,research papers).Be able
58、to manipulate and format content in order to better memorize,use and share it with other HCPs.SOLUTION Centralize qualitative medical content to make it searchable by a Generative AI model.Train the proper LLM(e.g.,Med-PalM by Google)to support identified prompts(i.e.,summarization,source identifica
59、tion,medical questions).Use a testing and validation framework,to assess summarization quality,question answering accuracy,recommendation relevance,etc.RESULTS Increased expertise,enabling healthcare professionals to effectively manage a wider range of medical conditions.I mproved efficiency,allowin
60、g healthcare professionals to spend more time focusing on patient care.Improved evidence-based decision-making.Healthcare professional(HCP)administrative assistantTARGETMedical practitioner:As a medical practitioner,I diagnose and treat medical conditions,conduct assessments,prescribe medications,an
61、d teach patients to improve their health and well-being while following ethical and medical guidelines.CONTEXTMedical practitioners are guided in their care decision-making by standards of care recognized by the global healthcare community and by recommendations6 issued by global or national healthc
62、are regulatory bodies,such as the Haute Autorit de Sant(HAS)7 in France.Over the years,regulatory recommendations have multiplied and become more complex.For patients with multiple pathologies,it can be time consuming for doctors to navigate these numerous standards and recommendations and apply the
63、m to patient care routines.CHALLENGE Support medical practitioners in care decision-making by generating evidence-based and personalized next step proposals.Bring care practices closer to standards of care and regulatory recommendations.SOLUTION Continuously collect all past and current patient jour
64、ney data.Train an LLM on a curated data set of global and national healthcare regulatory recommendations and key scientific publications describing current and emerging standards of care.Contextualize the LLM with patient historical data and ad hoc practitioner insights.Generate evidence-based,perso
65、nalized,explainable next steps in a diagnosis or care decision-making context with relevant sources available for review.2 APPLICATIONS&BENEFITS2 APPLICATIONS&BENEFITS1617GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCAREUSE CASE#4TARGETHealth information manager:As a health information manage
66、r,I collect,analyze,code,maintain,and secure a healthcare facilitys medical information for billing,research,quality improvement,and credentialing.CONTEXTMedical procedures require a code8 for transmission to insurance or social security systems.This code ensures accurate identification of the medic
67、al act for reimbursement purposes.Currently,coding is manually entered into systems,resulting in errors and omissions.From a business perspective,it also represents a loss of revenue and time for health information managers.CHALLENGE Automate the most manual aspects of medical procedure coding.Suppo
68、rt health information managers with billing code inquiries,insurance regulations,reimbursement guidelines,and other billing-related questions.SOLUTIONUse todays software tools that use AI and machine learning algorithms to extract relevant information from patient data,such as electronic health reco
69、rds(EHRs).Leverage these tools to generate accurate medical codes9 and thus improve interactivity.Once coding is generated and validated,automate the process of sending bills to social security and insurance companies.RESULTS Avoid medical coding mistakes.Avoid income loss for hospitals and clinics.
70、Medical coding assistant for hospitals and clinicsWe have not reached a stage where deploying Generative AI leads to job losses.While the technology holds immense potential in various fields,it often falls short in delivering results with the required precision.Vaibhav KULKARNIData Engineering LeadD
71、EBIOPHARMUSE CASE#5TARGETPatients:As a patient,I receive medical care,treatment,or attention from a healthcare provider,such as a doctor,nurse,or hospital.CONTEXTIn 2022,52%of people9 in the EU sought information on the Internet about their health or a medical procedure.And 70%10 of them were expose
72、d to misinformation without knowing it.CHALLENGE Enable patients to retrieve reliable information ahead of a surgery or a medical procedure.Educate patients on medical conditions,medications and preventive care.SOLUTION Train an LLM with curated data sets describing medical procedures,preparation ad
73、vice,medication posologies,adverse effects of medication,possible complications after surgery,and the main preventive reflexes to keep in mind by pathology.Set up a chatbot to provide patients with contextual and accurate responses to public health information11.Enhance patients knowledge about thei
74、r overall hospitalization journey12.Encourage patients to see a medical practitioner if the conversation reveals signs that care is needed.RESULTS Reduce stress linked to the hospitalization journey or any other medical treatment.Avoid misinformation by providing safe and private content,validated b
75、y health authorities.Preventive and informational agent for patients2 APPLICATIONS&BENEFITS2 APPLICATIONS&BENEFITS1819GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCARE1 Hyperscalers and tech providers3The healthcare ecosystem:Gearing up to unlock the potential of Generative AIGenerative AI is
76、 a rapidly evolving technology and the healthcare ecosystem has already begun its metamorphosis in an attempt to harness its full potential.To better understand the progress of this transformation,we can divide the ecosystem into four main groups of players:hyperscalers,startups,pharmaceutical compa
77、nies,and public domain actors.3 THE HEALTHCARE ECOSYSTEM2021GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCARE1.1 Hyperscalers democratize Generative AI by building accessible models and tools1.2 Hyperscalers have already begun to create healthcare domain-specific Generative AI models and serv
78、ices to better address the complexities of medical data Hyperscalers successfully launched generalist foundation models available to the public and companies,starting with ChatGPT.Since its public launch in November 2022,it has captured the worlds attention,showing millions of users around the globe
79、 the extraordinary potential of artificial intelligence.After their initial introduction to Generative AI,hyperscalers began taking on the challenge of fine-tuning these generalist models into domain-specific Large Language Models(LLMs).As healthcare is one of the areas with the greatest potential f
80、or transformation thanks to GenAI,it quickly became a top priority for hyperscalers.Like everyone else,the healthcare industry first discovered Generative AI through publicly available models,such as OpenAIs ChatGPT or Googles Bard.The ecosystem reacted with enthusiasm,as HCPs and pharmaceutical com
81、panies employees saw an opportunity to gain efficiencies in their everyday administrative operations.Beyond the most administrative healthcare industry use cases,hyperscalers have already started creating healthcare domain-specific Generative AI models and services to better meet medical data comple
82、xity requirements.There is a lot of excitement around these initiatives,and while first results are promising,HCPs and pharmaceutical companies should be cautious about using these models which are far from limitless(see part 3).We can look at four major examples:Med-Palm2 by Google 0102To enable th
83、ese administrative use cases,Big Tech companies developed applications such as Microsofts Copilot,which consists of embedding OpenAIs ChatGPT into Microsoft 365.However,the healthcare ecosystem also quickly raised concerns around the unbridled use of ChatGPT or Bard in a pharmaceutical or healthcare
84、 environment:healthcare data is sensitive,and should not be shared in the publicly available versions of LLMs(see part 3).To go beyond the basic augmented productivity use cases of LLMs and provide some reassurance with regard to data privacy,Big Tech created API services(GPT by OpenAI,PaLM by Googl
85、e)to allow companies to configure their own generalist LLM while maintaining control of their data.Most LLM use cases we see emerging in the healthcare industry are derived from these APIs.Companies either fine-tune or prompt engineer these generalist models to fit their specific needs.Example of US
86、MLE-style questionA 32-year-old woman comes to the physician because of fatigue,breast tenderness,increased urinary frequency,and intermittent nausea for 2 weeks.Her last menstrual period was 7 weeks ago.She has a history of a seizure disorder treated with carbamazepine.Physical examination shows no
87、 abnormalities.A urine pregnancy test is positive.The child is at greatest risk of developing which of the following complications?A.Renal dysplasiaB.MeningoceleC.Sensorineural hearing lossD.Vaginal clear cell carcinomaMicrosoft BioGPT“Apricitabine”“Janus kinase 3 JAK-3)”“Janus kinase 3(JAK-3)is a m
88、ember of the Janus kinase(JAK)family of non-receptor tyrosine kinases and plays an important role in the regulation of cell proliferation,differentiation,survival,migration,and angiogenesis.”Microsoft has created an LLM called BioGPT based on GPT architecture that is specifically intended for proces
89、sing text and biomedical data.It was trained on biomedical research articles so it can perform tasks such as answering questions,extracting relevant data,and generating text relevant to biomedical literature.For example,BioGPT can generate descriptions of a specific therapeutic class such as“Janus k
90、inase 3(JAK-3)or of a specific therapysuch as“Apricitabine.”BioGPT achieved 81%accuracy on PubMedQA.Med-PaLM is an LLM designed to provide high-quality answers to medical questions.It was created by retraining Googles generalist LLM models with data such as medical exams,medical research papers,and
91、healthcare consumer queries.Its first version was published in Nature in July 2023,and was the first AI to surpass the pass mark on US Medical License Exam(USMLE)-style questions.To date,Med-PaLM 2 achieves 86.5%accuracy on PubMedQA13,the dataset used by the healthcare community to benchmark LLM per
92、formance.“Apricitabine is an oral prodrug of 5-aza-2-deoxycytidine(5-aza-CdR),a DNA methyltransferase(DNMT)inhibitor,which has been approved by the US Food and Drug Administration(FDA)for the treatment of myelodysplastic syndrome(MDS)and acute myeloid leukemia(AML)in combination with low-dose cytara
93、bine(Ara-C)and granulocyte colony-stimulating factor(G-CSF)for patients with intermediate-2 or high-risk MDS or AML”INPUTBIOGPT OUTPUT3 THE HEALTHCARE ECOSYSTEM3 THE HEALTHCARE ECOSYSTEM2223GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCAREHealthScribe by AWSNVIDIA BioNeMo 0304In March 2023,Nv
94、idia launched a Generative AI platform for drug discovery that enables researchers to fine-tune Generative AI applications on their own proprietary data,and to run AI model inference directly in a web browser or through new cloud application programming interfaces(APIs).BioNeMo offers 9 open-source
95、foundation models with applications such as novel small molecule and protein sequence generation.HealthScribe is a clinical documentation service that enables healthcare software vendors to build clinical applications that use speech recognition and Generative AI to create transcripts of patient vis
96、its that auto-populate with additional relevant information,identify key details,and create summaries that can be entered into an electronic health record.It is currently being previewed for two specialties:general medicine and orthopedics.2 Startups3 THE HEALTHCARE ECOSYSTEM3 THE HEALTHCARE ECOSYST
97、EM2425GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCARE2.1 Healthcare startups play a complementary role to hyperscalers by providing the ecosystem with off-the-shelf innovative solutions to address their more specific problemsWhile Big Tech excels at delivering large-scale solutions that cut
98、 across all industries,startups have demonstrated the ability to develop solutions that address the specific needs of the healthcare industry.With support from the pharmaceutical industry,venture capital and private equity investors,multiple disruptive startups have emerged to develop Generative AI
99、applications around use cases that can deliver short-term value,such as HCP productivity,patient engagement and care decision making tools.To illustrate,we have selected three startups that offer off-the-shelf Generative AI modules:Nabla Copilot combines speech-to-text and Generative AI technologies
100、 to offer a digital assistant for HCPs that can automatically transcribe consultations in real time and generate summarized clinical notes.It works for both in-person and video consultations.One of the key differentiating factors of the solution is its rapid deployment,which is achieved through thre
101、e versions,all cloud-based:web-based app,Chrome extension,and API.In all cases,Copilot does not store or retain any data entered by the user.The API version can be integrated into the care centers electronic health record software.Based on publicly available information,Nabla is using OpenAIs GPT-3
102、LLM as the basis for Copilot,with the longer-term goal of building its own LLM,fine-tuned to the specific language and needs of medicine and healthcare.Memora Health offers AI-powered care enablement that helps clinicians focus on delivering high-quality care while proactively engaging patients alon
103、g complex care journeys.The platform relies on an LLM-enabled conversational AI capable of responding to routine patient messages.It operates on a retrieval-based approach,drawing on a selective but extensive clinician-validated database.This retrieval-based approach ensures that the AI only respond
104、s to patient questions with clinically-reviewed responses and only operates within clinician-validated pathways.More complex clinical concerns are automatically triaged to care teams.Memora HealthNablaHippocratic AI is building a safety-focused Large Language Model(LLM)designed specifically for heal
105、thcare,with an initial focus on non-diagnostic,patient-facing applications.This model is pre-trained on trusted,evidence-based healthcare content and already outperforms GPT-4 on 105 out of 114 healthcare exams and certifications.To ensure the safety of the model before it can be released to the pub
106、lic,the Hippocratic AI team is conducting a reinforcement learning with human feedback(RLHF)process using healthcare professionals to train and validate the models readiness for use.2.2 Investors are joining the Generative AI revolution and have already begun financing high-impact projectsPrivate eq
107、uity and venture capital firms are anticipating the impact of Generative AI in healthcare and have already started to invest in the technology.First,startups specializing in Generative AI are supported by an investment momentum strong enough to expand beyond the borders of their home country.Lets ta
108、ke another look at the French startup Nabla,which we mentioned earlier.Supported by Firstminute Capital and Artemis,they recently struck a deal with Permanente Medical Group in the United States,which is part of Kaiser Permanente,a healthcare giant with 75,000 employees.Nabla Copilot will initially
109、be rolled out in Northern California,but if the product proves effective,it will be implemented throughout Kaiser Permanentes presence in the United States.The challenge lies in integrating Generative AI into established companies that already have access to high-quality healthcare data,rather than
110、investing in new startups.Anne-Sophie Saint-Martin-Partner-NEWFUND-Seed VCHippocratic AI3 THE HEALTHCARE ECOSYSTEM3 THE HEALTHCARE ECOSYSTEM2627GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCARE3 Pharmaceutical companiesIn addition to funding new Generative AI startups,investors are also posit
111、ioning themselves to help established healthcare startups integrate this technology into their existing services.Ongoing projects focused on healthcare data processing,interoperability,or infrastructure are likely to explore Generative AI to augment their offerings.For example,Newfund,a French early
112、-stage venture capital firm,is backing Arkhn and Omnidoc,two healthtech startups in its portfolio,in their efforts to integrate Generative AI into their current services.Generative AI combined with quantum computing is opening new doors in drug discovery,and the transformative potential of this comb
113、ination is Generative AI is poised to generate numerous new opportunities in drug discovery.While most current projects are in the early stages of development,the combination of GenAI and quantum computing in drug discovery could not only lead to the creation of new treatments,but also to new advanc
114、es that nature itself is not yet able to offer.Florian Denis-Investment Director-ELAIAattracting significant interest from investors.This is because the probabilistic nature of quantum computing can enable Generative AI to explore a broader solution space,potentially discovering novel treatments,opt
115、imizing healthcare processes,and uncovering patterns previously beyond the reach of classical computing.Elaia and other leading investment firms recently backed Aqemia,a next-generation pharmatech company using quantum computing to generate one of the worlds fastest-growing drug discovery pipelines.
116、The disruptive speed and accuracy of their technology platform allows them to treat and scale drug discovery projects as technology projects.3 THE HEALTHCARE ECOSYSTEM3 THE HEALTHCARE ECOSYSTEM2829GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCARE3.1 Large pharmaceutical companies have already
117、 started to develop proofs of concept(POCs)using the large language models available on the marketTo ensure that the performance of AI models is maximized for all the patients we serve,we need to responsibly train these models on as many representative datasets as possible.When healthcare data is av
118、ailable at scale in a way that protects patient privacy,first and foremost we can create reliable databases and leverage the power of AI and machine learning to uncover insights that can help us better understand and measure disease onset and progression,design and optimize more targeted medicines,c
119、onduct more efficient and diverse clinical trials,and more.Tommaso Mansi-Vice President of AI/ML&Digital Health THE JANSSEN PHARMACEUTICAL COMPANIES OF JOHNSON&JOHNSONPharmaceutical companies have a long history of using AI.They see Generative AI as a way to broaden its applications and,most importa
120、ntly,to speed up the implementation of use cases and improve their ease of maintenance.To embark on an impactful Generative AI transformation and overcome the technical complexity inherent to medical data,most pharmaceutical companies would rather adopt a test-and-learn approach than rush into subst
121、antial investments towards industrialization.In fact,starting with an initial proof of concept enables them to quickly showcase concrete results and get people on board.This step is especially important as Generative AI is evolving on a daily basis,and pharmaceutical companies need to gain traction
122、through results for the entire field,not just a specific application or technology.The healthcare ecosystem,as well as both French and European regulators,are still learning about the potential and limitations of Generative AI.And the specificity of the healthcare domain is that most Generative AI u
123、se cases,especially those that are HCP and patient-facing,have a very low tolerance for error.Artefact already accompanies several pharmaceutical industry clients in the delivery of their first Generative AI POCs,for Research and Development,medical affairs,and marketing business departments.3.2 To
124、keep pace with Generative AI advances in drug discovery,most pharmaceutical companies chose to enter strategic partnerships with key medical AI startups3.3 A few examples of innovative collaborations to watchBefore the large language model hype,big pharmaceutical companies had already been experimen
125、ting with Generative AI in drug discovery and design for several years.To do so,most of them entered into strategic partnerships with key medical AI startups.Pharmaceutical companies invest in these innovative partnerships because traditional drug discovery and design has been an expensive and ineff
126、icient journey for too long.While it can take an average of 10 to 15 years to develop a drug,a 2023 article of the Drug Discovery Today magazine shows that a large pharmaceutical company spent an average of$4.4 billion annually on R&D and launched 0.78 new drugs between 2001 and 2020.Generative AI h
127、as already been shown to increase speed and efficiency at every stage of the process.It is being used to identify novel targets for disease and to design de novo molecules capable of acting on those targets.Some players are leveraging it to determine the likelihood of success in clinical trials.Insi
128、lico Medicine is a world-leading end-to-end Generative AI-driven biotech company with pipelines to explore transcriptome diversity by RNA-seq.They signed a strategic research collaboration with Sanofi in November 2022,worth up to$1.2 billion.Sanofi is leveraging their proprietary Pharma.AI platform
129、across biology,chemistry,and clinical development to accelerate the discovery of novel therapeutics.Aquemia is a next-gen pharmatech aimed at designing fast innovative drug candidates for critical diseases.They rely on unique quantum algorithms and Generative AI to design novel drug candidates.Servi
130、er entered a partnership with them in December 2021 to accelerate drug candidate discovery in immuno-oncology using artificial intelligence and recently announced the extension of their collaboration on a new undruggable target in the above mentioned therapeutic area.BenevolentAI is a clinical-stage
131、 AI and GenAI-enabled drug discovery company.Their key asset is the Benevolent Platform,a versatile,scalable and robust AI-enabled drug discovery platform built by expert scientists using multimodal data foundations.They also offer Generative AI products for knowledge exploration.AstraZeneca has bee
132、n their partner since 2019,and this collaboration is bearing fruit as,based on BenevolentAI technology,the pharmaceutical company is advancing four of the most promising targets selected for its portfolio in chronic kidney disease(CKD)and idiopathic pulmonary fibrosis(IPF),while good progress is bei
133、ng made on further target selections in heart failure and SLE.Iktos is an innovative company specializing in the development of artificial intelligence(AI)solutions applied to chemical research,specifically medicinal chemistry and new drug design.In March 2022,they announced the application of their
134、 Generative AI-driven de novo design software called Makya for de novo design to selected Pfizer small-molecule discovery programs.3 THE HEALTHCARE ECOSYSTEM3 THE HEALTHCARE ECOSYSTEM3031GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCARE4 Public domain4.1 Hospitals and research institutes have
135、 already begun to adapt this technology to their specific contextsWhile some practitioners remain skeptical about Generative AI,the revolution it is bringing has not passed healthcare by.The healthcare community quickly realized that the generalist models made available by Big Tech companies would n
136、ot be suitable for its needs.The use cases for Generative AI in healthcare have a much lower tolerance for errors compared to other industries.They also involve the processing of complex healthcare-specific queries,on which these models cannot perform well because they have not been trained on curat
137、ed healthcare data.In France,healthcare stakeholders have therefore begun to develop their own foundational models,with the support of public funding entities.To illustrate,we can look at the LLM4All Project,funded by the ANR(French National Research Agency)and led by public hospitals of Paris(AP-HP
138、),CNRS,Computer Science Laboratory of cole Polytechnique(LIX),and LINAGORA,an open source software and services company.In September 2023,the consortium announced its project aimed at providing retrained LLMs in open source to address some of the specific needs of the The AI revolution in healthcare
139、 is first and foremost a data revolution,not so much a modeling revolution.Emmanuel Bacry-Senior Researcher CNRS,Chair PRAIRIE,Chief Scientific Officer HEALTH DATA HUBhospital sector.These models will be validated through two specific use cases in French:automatic summaries of SAMU meetings,and anal
140、ysis of SAMU emergency calls.The call for projects,“Communs numriques pour lintelligence artificielle gnrative”,launched by the BPI in June 2023,will also be a major accelerator for Generative AI in healthcare in France.Several consortia of hospitals and medical research institutes have applied to d
141、evelop new sovereign,open source foundation models dedicated to healthcare.This call for projects launched by the BPI will also promote the creation of curated medical datasets for the purpose of training medical LLMs.The call for projects closed on October 24th,and while we do not yet have access t
142、o the full list of applications,we have learned through our panel of experts that key players such as the Health Data Hub,APHP,the University Hospital of Rennes(CHU Rennes)and the University Hospital of Reims(CHU Reims)have submitted large-scale applications in partnership with other hospitals and h
143、ealth research institutions.3 THE HEALTHCARE ECOSYSTEM3 THE HEALTHCARE ECOSYSTEM3233GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCARE4.2 Foundation LLMs provided by hyperscalers have one major limitation for the French healthcare ecosystem:they are not sovereign.In addition to emerging initia
144、tives by French hospitals and medical research teams,Docaposte,a French expert in sensible computing,recently announced the launch of its first sovereign LLM service with healthcare use cases.In partnership with three French companies(LightOn,Aleia,and NumSpot),Docaposte is launching its first sover
145、eign,cloud-based LLM,available from November 2023.This offering is aimed at French organizations that handle sensitive data,particularly in the healthcare sector.Docaposte unveiled a first application called“MyAssistant Sant”.They trained the foundation LLM called Alfred,provided by LightOn,using th
146、e reference documents that govern the patients medical record.The resulting model powers a chatbot that assists the doctor in gathering information about a patients case before treatment and more effectively transcribes information back into the same patient record.Generative AI in healthcare:Limita
147、tions,challenges and opportunities4To ensure the sovereignty of France and Europe in the field of Generative AI,public actors have started to equip themselves with supercomputers.Computing power is a limiting factor in the training of LLMs,and as these models are becoming larger and larger,supercomp
148、uters will be required to train them.Supercomputers are critical to a countrys sovereignty in artificial intelligence,and even more so in Generative AI.They enable advanced AI research and model training,which is essential to remain competitive globally.One of Frances main research organizations,the
149、 CEA,recently announced that it will host the future European Exascale supercomputer as of 2025,which will be one of the fastest supercomputers in the world.This supercomputer will bring Europe back into the supercomputing race and will act as a sovereign accelerator for several strategic European d
150、ata-related challenges,such as training the next generation of Generative AI or multimodal models.Although Generative AI holds the promise of revolutionizing the healthcare industry,it comes with its own set of risks and challenges that need to be addressed.3 THE HEALTHCARE ECOSYSTEM3435GENERATIVE A
151、I FOR HEALTHCAREGENERATIVE AI FOR HEALTHCAREGiven the highly sensitive nature of healthcare data in terms of confidentiality,it is inherently subject to a variety of privacy regulations and requirements.As artificial intelligence and machine learning become more integrated into the field of medicine
152、,there is a growing concern that inaccuracies in algorithms could lead to patient injury and medical liability.While certain regulations have already been put in place to protect patients personal data,such as the General Data Protection Regulation(GDPR)in Europe and the Health Insurance Portability
153、 and Accountability Act(HIPAA)in the United States,both national and international regulatory bodies are actively working to keep pace with the rapid advances in Generative AI.One notable example is the European Unions proposed AI Act,the worlds first concrete initiative to ensure that AI systems de
154、ployed in the EU are safe,transparent,1 The need for robust legal and regulatory frameworks to safeguard health data privacy,security,and integritytraceable,non-discriminatory,and environmentally friendly.It proposes a principle of graduated rules by classifying AI models according to the risk they
155、pose to users(minimal,limited,high or unacceptable risk),with associated obligations and counterbalancing measures(from minimal transparency requirements to impact assessment studies and measures to mitigate potential risks).While the primary objective is to safeguard the interests of citizens again
156、st unethical use of their data,a commendable goal,it inadvertently poses challenges to data access and sharing.The regulations aim to strike a balance between privacy and technological advancement.Will Europe lag behind other regions in the development and adoption of Generative AI,or will it set a
157、precedent for future global rules and standards in this rapidly evolving field?It is too early to predict what this regulation will mean for the future.2 Ensuring privacy and safety around healthcare data and Gen AI models underscores the need for innovative toolsWhile data anonymization is the most
158、 common approach for safeguarding privacy,it can be a complex task for medical data sources.The main challenge does not lie in the removal of overtly personal information such as names or social security numbers,which can be readily redacted or randomized.The true complexity emerges when addressing
159、indirect data points,such as social connections(often referenced in medical reports),intricate medical histories,or nuanced spoken language.While these elements may not be distinctly identifying on their own,their collective amalgamation can yield powerful identifiers capable of describing a unique
160、individual or an exceedingly small cohort.Legal risksConsidering the ongoing uncertainty surrounding the full extent of this technologys implications,it is crucial for industry stakeholders and users to engage in meticulous deliberation,especially when addressing the legal aspects in the healthcare
161、sector.These legal aspects include issues such as liability for medical errors,limitations on data flow due to privacy and confidentiality regulations,and the crucial question of data sovereignty for data owners.Data and technologyFrom a technological point of view,Gen AI faces several key challenge
162、s.On one hand,it relies heavily on deep learning techniques to produce new content,leading to a certain opacity in its decision-making process and a risk of biases emerging.On the other hand,its effectiveness depends on the availability of a large amount of data which,in the medical field,is often d
163、ifficult to acquire due to privacy concerns and data access restrictions,therefore limiting its capabilities at present.Training and adoptionIt is imperative for institutions and companies to implement well-structured change management processes in order to integrate AI into their operations.Equally
164、 crucial is the need for comprehensive education and awareness programs for healthcare professionals and C-suite executives to provide them with a deep understanding of both the capabilities and limitations of AI,as well as an awareness of its vulnerabilities.This knowledge is essential to effective
165、ly mitigate and balance the inherent risks associated with the use of AI in healthcare.However,these challenges also open up new opportunities for disruptive players to leverage their existing assets and expand their offerings into the realm of Generative AI.This section,based on the interviews we c
166、onducted and our experience in delivering client projects,highlights the most significant risks related to Generative AI in healthcare.It also explores potential strategies for mitigating these risks to ensure responsible and ethical implementation,along with the opportunities that arise.CHALLENGES4
167、 LIMITATIONS,CHALLENGES&OPPORTUNITIES4 LIMITATIONS,CHALLENGES&OPPORTUNITIES3637GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCARE3 Sarus,an innovative start-up addressing privacy issues4 At present,the ability to explain the decision-making process of an AI remains elusive,making the explainab
168、ility of LLM models a significant challengeEmbracing AI is not just about integrating the technology,but also understanding its decisions while mitigating biases.Algorithms are often referred to as“black boxes”because of their ability to evolve as new data is integrated and because they rely on deep
169、 learning algorithms to create new content.In addition,there is growing evidence that these algorithms can sometimes produce recommendations that are racially or otherwise biased.Bias awareness is crucial,as most users struggle to distinguish between truth and the likelihood of a result.While truth
170、stands as an unambiguous and factual representation-essential for a definitive medical diagnosis or assessment-likelihood operates within the realm of probabilities.Indeed,AI generally relies on statistical inference and its output is determined by what is“most likely”,given its training data,algori
171、thms,To circumvent the challenges of anonymization,Sarus offers LLM fine-tuning powered by differential privacy.It ensures that no personal information is embedded in the fine-tuned LLM.In short,Differential Privacy(DP)is a mathematical approach that protects personal information during data process
172、ing.This goal is achieved by adding controlled noise to the computations,making it virtually impossible to identify individual data points.This noise is adjusted to protect privacy while still allowing for valuable statistical analysis.This methodology protects patient data without the need for anon
173、ymization.When prompted about some topics they dont know,LLMs tend to“hallucinate”wrong responses.One way to mitigate this phenomenon is to automatically retrieve the documents in a knowledge base that are most likely to contain elements of the response and add them to the prompt so that the LLM has
174、 more context to give a correct answer.This technique is called Retrieval Augmented Generation(RAG).Other techniques can be used depending on the use case,such as:fine-tuning.When dealing with confidential data,these approaches should be combined with techniques to prevent the leakage of confidentia
175、l data in LLM responses.Nicolas Grislain-Co-founder and Chief Scientific Officer-SARUSand the nature of queries.In particular,LLMs often generate responses that may appear authoritative and plausible to end users,yet they may be completely incorrect or contain serious errors.This phenomenon known as
176、“hallucination”is a common bias inherent in LLMs.In a sector such as healthcare,it is imperative that this problem is addressed at its source,as it could pose a risk to patients.While some solutions(such as Retrieval Augmented Generation(RAG),fine-tuning LLMs)already exist,their widespread implement
177、ation remains a work in progress.In fact,the World Health Organization(WHO),which published the Ethics&Governance of Artificial Intelligence for Health in 2021,recently called for caution because“the data used to train AI may be biased,generating misleading or inaccurate information that could pose
178、risks to health,equity,and inclusiveness.”5 LLMon,an innovative approach for detecting AI Safety risks in LLM applicationsThe solution,developed by French startup Giskard,aims to provide a robust defense against potential pitfalls such as hallucinations,ethical biases,and inaccuracies that can occur
179、 in LLM-based applications.By verifying that outputs are accurate,it is designed to be a steadfast companion,offering comprehensive monitoring and evaluation capabilities for LLM-driven endeavors.Even with the development of AI security-enhancing tools,it remains crucial for both professionals and p
180、atients to understand the difference between having confidence in an answer and granting unconditional trust.This nuanced understanding serves as a catalyst for more informed decision-making and acts as a safeguard against overreliance on AI,emphasizing the essential role of human oversight.4 LIMITA
181、TIONS,CHALLENGES&OPPORTUNITIES4 LIMITATIONS,CHALLENGES&OPPORTUNITIES3839GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCARE6 In addition to technology,data accessibility plays a pivotal role in providing the necessary input for the operation of Generative AI models7 The ecosystem must find crea
182、tive solutions to address data scarcity while respecting privacy and regulationsThe significant progress made by Generative AI depends primarily on the accessibility of the data needed to fuel these emerging models.Currently,a major challenge is the collection of large amounts of high-quality data t
183、o ensure reliable outcomes.Even in an era characterized by data ubiquity,obtaining top-quality data in the healthcare sector faces various technical and ideological obstacles.The scarcity of healthcare data can hinder the development and training of Generative AI models,which rely on large and diver
184、se datasets to learn effectively and produce accurate and reliable outcomes.Another phenomenon exacerbating the challenge of data accessibility is the distinct concern about data monetization that has emerged,particularly in the French landscape.While the trend is moving towards a greater willingnes
185、s to share data,the difficulty of evaluating the financial value of the services associated with data and/or databases was further accentuated during the COVID-19 pandemic,when data sharing was needed more than ever.Paradoxically,this period witnessed the reinforcement of a“data ownership”fantasy,wh
186、ich ran counter to the expectation that it would diminish precisely when the demand for collective data sharing was at its peak.Key stakeholders in the healthcare ecosystem who have not yet fully embraced Generative AI bear a significant responsibility in providing high-quality medical data for trai
187、ning Large Language Models(LLMs),such as medical software editors like Cegedim,Doctolib or Osiris.Moreover,healthcare-centric AI companies like Owkin,are already collaborating with numerous hospitals and research institutions worldwide through their federated research network.Medical data warehouse
188、providers like Arkhn or Codoc also have the potential to leverage their network of hospital partners to provide both structured and unstructured data for training LLMs.As access to healthcare data is particularly difficult,one of the key differentiators for organizations wishing to embark on a Gener
189、ative AI journey will be their ability to access large quantities of data.Benjamin Belot-Partner-KURMA PARTNERS,VC focusing on Healthtech&BiotechIn addition,another way to overcome the scarcity of data challenge is to rely on GenAIs ability to generate synthetic data,such as medical images.This appr
190、oach allows training data to be augmented,thereby promoting diversity in medical research and training.It mitigates privacy concerns by removing sensitive patient information while retaining the essential features needed for meaningful analysis.It thus provides an answer to the common problem of dat
191、a scarcity and confidentiality,facilitates data sharing,and improves model robustness.8 Data acculturation and training are indispensable to familiarize healthcare professionals with the use and potential associated risks of this technologyOver the long term,humans risk becoming overly reliant on ge
192、nerated documents,which could lead to a loss of comprehension and technical skills.Human decision-making must be preserved to prevent alienation caused by LLMs.Vincent Vuiblet-Professor of Universities and Hospital Practitioner-CHU REIMS,URCA and Director of lInstitut dIntelligence Artificielle en S
193、ant Reims Champagne Ardenne(I2AS)It is essential to frame the use of a Generative AI model with a usage convention and to ensure in-depth training for health professionals regarding its intrinsic constraints and vulnerabilities.Jean-Marc Bereder-Artificial Intelligence Usage Specialist and Former He
194、ad of department-NICE UNIVERSITY HOSPITALA lack of public awareness of biases in data,models,and their applications can lead to potential misinterpretation of the tool,its scope,and its results.Institutions and companies should prioritize the implementation of comprehensive change management process
195、es for the seamless integration of AI.This entails more than just installing software;it involves a fundamental reevaluation of the organizations core values.It includes educating collaborators about AIs capabilities and constraints,while highlighting its pivotal role in achieving the organizations
196、objectives,such as enhancing patient care and providing time-saving tools.To ensure the genuine effectiveness of AI,including GenAI,it is imperative not only for the technology to be state-of-the-art but also for the professionals using it to be highly skilled.This is especially critical for healthc
197、are professionals who are primary users and directly interact with patients.Currently,there is a notable lack of comprehensive training programs that are tailored to different user proficiency levels,from novice to expert.Such programs are essential for making professionals not only operationally pr
198、oficient,but also culturally sensitive to the nuanced aspects of AI.4 LIMITATIONS,CHALLENGES&OPPORTUNITIES4 LIMITATIONS,CHALLENGES&OPPORTUNITIES4041GENERATIVE AI FOR HEALTHCAREGENERATIVE AI FOR HEALTHCAREGenerative AI is a game changer for the healthcare industry,transforming the way we interact wit
199、h data,and consequently how we analyze,make decisions,and personalize our interactions.This year,we have observed an extremely rapid increase in the awareness among all players in the ecosystem:tech players are developing specific enablers to better address healthcare challenges,while research insti
200、tutions,industry players,caregivers,and public institutions have mostly initiated consultations and initial projects.However,this enthusiasm should be tempered as the technology is still too new and immature.Many players are moving forward knowing that the potential is significant,but with a high de
201、gree of uncertainty about the achievable performance,the possible level of industrialization,and the regulatory constraints that will be defined.Trust and control play a critical role in realizing the potential of Generative AI in healthcare.It should therefore be seen as a human transformation,not
202、just a technical one.5To overcome these challenges,it is essential that healthcare and pharmaceutical industry stakeholders keep humans at the center of the Generative AI transformation at all times.First,humans should be at the center of decision-making processes.Individuals should not only be trai
203、ned in the use of these technologies but also equipped with the tools and capabilities to exercise control and make informed decisions.Indeed,the primary objective of AI in healthcare is to empower the medical ecosystem,ensuring that technology serves as an enabler of human expertise rather than a r
204、eplacement for it.Second,we must ensure that humans remain the primary beneficiaries of this transformation.Indeed,the widespread diffusion and adoption of these technologies,if perceived only as productivity tools,could lead to a very technological view of health,thus depersonalizing the patient ca
205、re experience by perceiving patients only as“data sets”inputs and outputs for models.However,the productivity gains of Generative AI applications should first be seen as a means to improve the overall quality of healthcare and patient management.Ultimately,the responsible and sustainable implementat
206、ion of Generative AI in the healthcare sector hinges on a delicate balance:harnessing AI for its capabilities while maintaining strong human oversight to ensure ethical and responsible use.Building and thinking as an ecosystem with common challenges and complementary strengths,is more necessary than
207、 ever for healthcare and pharmaceutical stakeholders to realize the full potential of data,AI,and Generative AI in the service of patients.5 CONCLUSION5 CONCLUSIONThanks&acknowledgementsWe also extend our heartfelt thanks to each member of our team for their tireless efforts and invaluable contribut
208、ions throughout the entire process.Pr Stphanie Allassonnire-Professor and Vice-President Valorisation and industrial partnerships,Chair PRAIRIE-Universit Paris CitEmmanuel Bacry-Senior researcher CNRS,Chair PRAIRIE-Chief Scientific Officer-Health Data HubBenjamin Belot-Partner at Kurma Partners-VC f
209、ocusing-Healthtech&BiotechJean-Marc Bereder-Artificial Intelligence Usage Specialist and Former Head of department-Nice University HospitalMarguerite Brac de la Perrire-Partner Attorney at Law in Digital Health-FieldfisherThomas Clozel-CEO and Co-founder-OwkinFlorian Denis-Investment Director-ElaiaN
210、icolas Grislain-Co-founder and Chief Scientific Officer-SarusVaibhav Kulkarni-Data Engineering Lead-DebiopharmTommaso Mansi-VP of AI/ML&Digital Health-The Janssen Pharmaceutical Companies of Johnson&JohnsonDr Grgoire Pign-CEO-PulseLife-Oncologist and Radiation TherapistAnne-Sophie Saint-Martin-Partn
211、er-Newfund-Seed VCPr Eric Vibert-MD,PhD,IKO,Liver Surgeon in Paul Brousse Hospital AP-HP,Paris Saclay University-Chairman of BOPAPr Vincent Vuiblet-Professor of Universities and Hospital Practitioner-CHU Reims and URCA-Director of Institut dIntelligence Artificielle en Sant Reims Champagne Ardenne(I
212、2AS)Artefact editorial team:Paul de Balincourt-Director-Healthcare Practice-ARTEFACT Tanguy Masgnaux-Senior Consultant-Healthcare Practice-ARTEFACTSbastien Marguers-Programs&Scientific Manager-AI for HealthLa Giroulet-Senior Consultant-Healthcare Practice-ARTEFACTEmma Tordo-Consultant-Healthcare Pra
213、ctice-ARTEFACTArtefact sponsors of the initiative:Justine Nerce-Managing Partner-Healthcare Practice Lead-ARTEFACTDamien Gromier-Founder&CEO-AI for HealthStphanie Trang-General Manager-AI for Life01 Clinical trials phases description:Phase I evaluates the tolerance of a human body after the administ
214、ration of the drug.Phase II evaluates the efficiency of the drug.Phase III ensures at a larger scale the efficiency and safety of drugs.02 “The minimum threshold is usually set to 300”-Stphanie Allassonnire,202303 The retention of patients in the whole clinical trial duration can change for various
215、reasons(distance,understanding,health complications,.),leading to some data gaps or lack throughout the process.04 Patient journey data(not exhaustive):it includes past surgeries,comorbidities,medication and duration of treatment as well as the age,gender,height,and weight of the patient.05 Syntheti
216、c Data Generation by Artificial Intelligence to Accelerate Research and Precision Medicine in Hematology06 HAS website07 Purpose of these Institutions,through the standards,is to guarantee the quality and safety of care for patients08 Implement Gen AI in medical acts coding09 Eurostat10 AHIMA Founda
217、tion:Spotting health misinformation online11 Example of the Covid-19 chatbot.12 Example addressed by Pr.Vincent Vuiblet:patients will be able to increase their knowledge about their hospitalization,to understand each step(before,during,and after the hospitalization).13 MultiMedQA is a comprehensive
218、collection of multiple-choice medical question-answering datasets,used for training and evaluating Med-PaLM.MultiMedQA is comprised of the following datasets:MedQA,MedMCQA,and PubMedQA.PubMedQA is usually the data set used for evaluation.We would like to express our sincere gratitude to all the inte
219、rviewees from all parts of the ecosystem(researchers,hospital practitioners,startups,pharmaceutical companies,investors,attorney)who contributed to the creation of this white paper.Their expertise and collaborative spirit were instrumental in shaping and refining our ideas.WE OFFER END-TO-END DATA&A
220、I SERVICES42GENERATIVE AI FOR HEALTHCARESPORTS&ENTERTAINMENT TRAVEL&TOURISM PUBLIC&GOVERNMENT REAL ESTATE MANUFACTURING&UTILITIESFMCG RETAIL&ECOMMERCE LUXURY&COSMETICS HEALTHCARE BANKING&INSURANCE TELECOMMUNICATIONSGlossaryArtificial Intelligence(AI):The field of computer science that focuses on cre
221、ating intelligent machines capable of performing tasks that typically require human intelligence,such as visual perception,speech recognition,problem solving,and decision-making.Machine learning:A subset of AI that involves developing algorithms and models capable of learning patterns and making pre
222、dictions or taking actions based on data,without being explicitly programmed.Deep learning:A subset of machine learning that utilizes artificial neural networks with multiple layers to process and learn from large amounts of data,enabling the model to make complex and sophisticated predictions or de
223、cisions.Foundation model:A pre-trained and highly capable AI model that serves as a basis or starting point for developing more specialized models,enabling faster development and reduced training time for specific tasks or domains.Generative AI:A branch of artificial intelligence(AI)that focuses on
224、creating and generating new content,such as images,music,or text,using algorithms and models.Large Language Model(LLM):A powerful AI model capable of processing and generating human-like text,leveraging a vast amount of pre-existing language data to generate coherent and contextually relevant respon
225、ses.AI Act:Proposed regulation by the European Union that aims to be the worlds first concrete initiative for regulating AI by establishing a legal framework for the development,deployment,and use of AI systems.AI bias:A phenomenon that occurs when an AI algorithm produces results that are systemica
226、lly prejudiced due to erroneous assumptions in the machine learning process.Clinical trial:A type of research that studies new tests and treatments and evaluates their effects on human health outcomes.Cohort:A group of people who share a particular characteristic,such as social and health factors.Di
227、fferential privacy:A mathematical technique that adds a controlled amount of randomness to a dataset to prevent anyone from obtaining information about individuals in the dataset.General Data Protection Regulation(GDPR):A comprehensive data protection regulation in the European Union(EU)that governs
228、 the collection,processing,and storage of personal data,and aims to protect the privacy and rights of individuals.Healthcare Professional(HCP):Any natural person who is a member of the medical,dental,pharmacy or nursing professions,or any other person who,in the course of his or her professional act
229、ivities,may prescribe,purchase,supply,recommend or administer a medicinal productHyperscalers:Large cloud service providers that offer services such as massive computing resources and storage at scale(AWS,Microsoft Azure,Google Cloud Platform)Model fine-tuning:The process of adjusting and optimizing
230、 a pre-trained AI model by exposing it to additional data,specific to the desired task or domain,in order to enhance its performance and make it more suitable for the target application.Programming language:A formal language with a set of rules and syntax used to write computer programs.It provides
231、the necessary instructions to tell a computer what tasks to perform,and it can be used to develop applications,algorithms and AI models.Prompt engineering:The process of designing and optimizing prompts or instructions given to a language model with the goal of eliciting specific and desired outputs
232、 or responses from the model.Real-world evidence(RWE):Clinical evidence of the safety and efficacy of a medical device generated using real-world data(RWD)from routine health care delivery.Retrieval Augmented Generation(RAG):An AI framework for improving the quality of LLM-generated responses by gro
233、unding the model on external sources of knowledge to supplement the LLMs internal representation of information.Synthetic data:Information thats artificially created rather than generated by real-world events to augment or improve AI models.Tech providers:Company or organization that offers various
234、technology solutions,products,or services for specific aspect(s)of an industry.CONTACT+33 1 79 72 45 HEADQUARTERS19,rue Richer75009 Paris FranceABOUT ARTEFACTArtefact is a global leader in data&AI consulting and data-driven marketing services,dedicated to transforming data into business impact and t
235、angible results across the entire value chain of organizations.Artefacts skyrocketing growth is fueled by our visionary and entrepreneurial founders as well as a unique methodology and multidisciplinary teams.Artefact helps clients put consumers at the heart of their strategy through a comprehensive
236、 range of data-driven solutions.Designed to meet the specific needs of todays clients,our services are delivered via a business-centric approach that is built upon deep AI expertise.We are a connected independent global network and we partner with 1000+ambitious clients around the globe.Our 1300+emp
237、loyees,present in 18 countries(Europe,the Americas,Asia,Middle East Africa)are focused on accelerating digital transformation thanks to a unique mix of company assets:cutting-edge AI technologies,agile methodologies for fast delivery and efficient scalability,and teams of market-leading experts in data science and digital marketing,always working together and focusing on business innovation.IN PARTNERSHIP WITH