Présentation Marc Cuggia
Présentation Marc Cuggia
Applications,
enjeux et verrous
Pr Marc CUGGIA
PUPH Informatique Médicale
Inserm LTSI - équipe Domasia
Centre de Données Clinique – CHU de Rennes & Université de Rennes
L’histoire Syndrome de corde attachée
de
Courtney
et Alex
3000,00
2500,00
Explosion de la 2000,00
recherche de 1500,00
500,00
0,00
70 72 74 76 78 80 82 84 86 88 90 92 94 96 98 00 02 04 06 08 10 12 14 16 18 20 22 24
19 19 19 19 19 19 19 19 19 19 19 19 19 19 19 20 20 20 20 20 20 20 20 20 20 20 20 20
Performances des LLM pour réussir
l’examen de médecine aux US (USLME)
Seuil de
réussite
Explosion des
performances
LLM+ GPU + Cloud
Mittal, A., 2023. Revolutionizing Healthcare: Exploring the Impact and Future of
Large Language Models in Medicine [WWW Document]. [Link]. URL
[Link]
of-large-language-models-in-medicine/ (accessed 8.12.24).
Au programme de ce topo
.
Partie I :
L'IA révolutionne la médecine :
Accélération de la
Personnalisation des Optimisation des recherche : découverte
Amélioration des diagnostics : ressources :
traitements : médecine de de médicaments,
plus précis, plus rapides, plus automatisation des
précision, thérapies identification de cibles
précoces. tâches, allocation des
ciblées, adaptation aux thérapeutiques,
besoins individuels. ressources, réduction meilleure
des coûts. compréhension des
maladies.
IA en action :
Détection des
cancers de la peau • Beltrami, Eric J., Alistair C. Brown, Paul J. M. Salmon,
David J. Leffell, Justin M. Ko, et Jane M. Grant-Kels. « Artificial
Intelligence in the Detection of Skin Cancer ». Journal of the
American Academy of Dermatology 87, no 6 (1 décembre
2022): 1336-42. [Link]
IA en action :
Identification de patients à risque de sepsis en soins intensifs
Kallonen, A., Juutinen, M., Värri, A., Carrault, G., Pladys, P., Beuchée, A., 2024. Early detection
of late-onset neonatal sepsis from noninvasive biosignals using deep learning: A multicenter
prospective development and validation study. Int J Med Inform 184, 105366.
[Link]
IA en action
Création de prothèses • MIT Media Lab. « Project Overview ‹ Neural Control of Upper Extremity
Prostheses ». Consulté le 7 août 2024.
intelligentes contrôlées [Link]
prostheses/overview/.
par la pensée
[Link]
• Focus sur l'imagerie médicale :
Détection automatique de lesions :
Fractures aux urgence (CHU de Rennes)
Le Quotidien du Médecin. « À Rennes, l’IA au secours des urgences ». Consulté le 7 août 2024.
[Link]
urgences.
Planification
interventionelle
en radiothérapie
• Segmentation d’organe
• Simulation de délivrance de dose
• Personnalisation du traitement
Texier, Blanche, Cédric Hémon, Adélie Queffélec, Jason
Dowling, Igor Bessieres, Peter Greer, Oscar Acosta, et al. « 3D
Unsupervised deep learning method for magnetic resonance
imaging-to-computed tomography synthesis in prostate
radiotherapy ». Physics and Imaging in Radiation Oncology 31 (1
juillet 2024): 100612.
[Link]
• Essai prospectif et populationnel en Suède su 80.000 patientes
• l'IA a été mise en œuvre en tant que lecteur indépendant de mammographies
• parallèlement à deux radiologues
• pour comparer les stratégies et leur impact sur les taux de détection du cancer.
Dembrower, Karin, Alessio Crippa, Eugenia Colón, Martin Eklund, et Fredrik Strand. « Artificial
Intelligence for Breast Cancer Detection in Screening Mammography in Sweden: A Prospective,
Population-Based, Paired-Reader, Non-Inferiority Study ». The Lancet Digital Health 5, no 10
(octobre 2023): e703-11. [Link]
• Extraction de features
correlées avec les
estimations du cancer
• Prédiction de 65% de la
variation de la
prévalence du cancer du
poumon à Los Angeles
• Forte corrélation avec
zone géographique en
défaveur socio-
économique
Harrer, Stefan, Pratik Shah, Bhavna Antony, et Jianying Hu. « Artificial Intelligence for Clinical Trial
Design ». Trends in Pharmacological Sciences 40, no 8 (1 août 2019): 577-91.
[Link]
L'IA et la recherche médicale :
Recrutement des patients éligibles
INCLUSION CRITERIA:
1. Inclusion criteria are same in both Phase I and Phase II parts, except for the number of prior disease relapses
2. Patients must have pathologic diagnosis of anaplastic astrocytoma defined as World Health Organization (WHO) grade III or
glioblastoma/gliosarcoma, WHO grade IV, which are confirmed by National Cancer Institute (NCI) Laboratory of Pathology. If the
pathology diagnosis is anaplastic glioma or anaplastic oligoastrocytoma, evidence of either intact combined loss of the short arm
chromosome 1 (i.e. 1p) and the long arm of chromosome 19 (1p/19q) chromosomes or molecular features suggesting astrocytic tumor
must be present (including, but not limited to alpha-thalassemia/mental retardation, alpha-thalassemia/mental retardation, X-linked
(ATRX), tumor protein P53 (TP53).
3. Patients must have recurrent disease, histologically proven or imaging suggestive of recurrent disease as determined by principal
investigator (PI). Prior implantation of Gliadel wafers is acceptable, if tumor recurrence is confirmed by histologic examination of the
recurrent tumor
4. Patients must have the ability to understand and the willingness to sign a written informed consent document.
5. Patients must be greater than or equal to 18 years old.
6. No more than two prior disease relapses to be eligible for the phase I portion of the study and no more than one prior relapse to be
eligible for phase II.
7. Patients must have undergone prior standard therapy for their primary disease. For patients with glioblastoma, this would include
surgical resection, or biopsy, if safe resection was not permitted due to the tumor location, radiation and adjuvant temozolomide. For
patients with anaplastic astrocytoma, this would include surgical resection, radiation and adjuvant chemotherapy procarbazine,
lomustine (CCNU) and vincristine (PCV) or temozolomide.
8. Tumor tissue must be available for review to confirm histological diagnosis.
9. Tumor block or unstained slides must be available for molecular profiling.
10. Karnofsky > 60 percent
11. Patients must have adequate bone marrow function (absolute neutrophil count (ANC) > 1,500/mm^3, platelet count of >
100,000/mm^3), adequate liver function (alanine aminotransferase (ALT) and aspartate aminotransferase (AST)< 3 times upper limit
normal and alkaline phosphatase < 2 times upper limit normal, total bilirubin < 1.5mg/dl), and adequate renal function (blood urea
nitrogen (BUN) < 1.5 times institutional normal and serum creatinine < 1.5 mg/dl) prior to registration. These tests must be performed
within 14 days prior to registration. Total bilirubin: patients with Gilbert's Syndrome are eligible for the study. (Total bilirubin level can
be exempted from the eligibility criterion.)
12. Patients must have recovered from the toxic effects of prior therapy to less than grade 2 toxicity per Common Terminology Criteria
(CTC) version 4 (except deep vein thrombosis)
13. At the time of registration, subject must be removed from prior therapy as follows:
14. greater than or equal to (28 days) from any investigational agent,
15. greater than or equal to 4 weeks (28 days) from prior cytotoxic therapy,
16. greater than or equal to 2 weeks (14 days) from vincristine,
17. greater than or equal to 6 weeks (42 days) from nitrosoureas,
18. greater than or equal to 3 weeks (21 days) from procarbazine administration,
19. greater than or equal to 1 week (7 days) for non-cytotoxic agents, e.g., interferon, tamoxifen, thalidomide, cis-retinoic acid, etc.
radiosensitizer does not count.
20. Patients having undergone recent resection of recurrent or progressive tumor will be eligible given all of the following conditions apply:
21. At least 2 weeks (14 days) have elapsed from the date of surgery and the patients have recovered from the effects of surgery.
EXCLUSION CRITERIA:
1. Patients who are receiving any other investigational agents. However, prior enrollment on a study using investigational agents is
acceptable
2. Patients with prior bevacizumab use for tumor treatment. Patients who received bevacizumab for symptom management, including but
not limited to cerebral edema, pseudoprogression can be included in the study (To date, there have been no effective regimens
developed for recurrent malignant gliomas that are refractory to bevacizumab. Inclusion of this patient population may impact the
ability to determine the efficacy of Zotiraciclib (TG02) with Temozolomide (TMZ.)
3. Any serious medical condition, laboratory abnormality, or psychiatric illness that would prevent the subject from providing informed
consent.
4. Any condition, including the presence of clinically significant laboratory abnormalities, which places the patient at unacceptable risk if
he/she were to participate in the study or confounds the ability to interpret data from the study. These would include:
5. Active infection (including persistent fever) including known history of human immunodeficiency virus (HIV) or Hepatitis C infection,
because these patients are at increased risk of lethal infections when treated with marrow-suppressive therapy.
6. Diseases or conditions that obscure toxicity or dangerously alter drug metabolism
7. Serious concurrent medical illness e.g. symptomatic congestive heart failure
8. History of allergic reactions attributed to compounds of similar chemical or biologic composition to temozolomide and/or Zotiraciclib
10.
of 3 years, are ineligible.
Zotiraciclib (TG02) is primarily metabolized by Cytochrome P450 1A2 (CYP1A2) and Cytochrome P450 3A4 (CYP3A4). Patients automatic systems for recruitment of patients to
receiving any medications or substances that are strong inhibitors or inducers of CYP1A2 and/or CYP3A4 are ineligible.
11. Patients, who continue to have prolonged corrected QT interval (QTc) (males: greater than 450ms; females: greater than 470ms as
calculated by Fridericia s correction formula) despite normal electrolyte balance and discontinuation of medications known to prolong
clinical trials. Int J Med Inform. 2011 Jun;80(6):371-88.
doi: 10.1016/[Link].2011.02.003. Epub 2011 Apr 2.
QTc, will be excluded from the study.
PMID: 21459664.
L'IA et la recherche médicale :
Recrutement des patients éligibles
L'IA et la recherche médicale :
Apport des jumeaux numériques
Potentiel : Sélection de cohorte, Bras synthétique dans les essais, essai à sujet unique, drug repurposing
L'IA et la recherche médicale :
Apport des jumeaux numériques pour les essais in sillico
Guevara M, Chen S, Thomas S, Chaunzwa TL, Franco I, Kann BH, Moningi S, Qian JM, Goldstein M, Harper S, Aerts
HJWL, Catalano PJ, Savova GK, Mak RH, Bitterman DS. Large language models to identify social determinants of
health in electronic health records. NPJ Digit Med. 2024 Jan 11;7(1):6. doi: 10.1038/s41746-023-00970-0. PMID:
38200151; PMCID: PMC10781957.
L'IA et la recherche médicale :
Suivis des patients dans les essais via un tchabot
• L’engagement et l’éducation des patients et de leurs familles, ont un impact sur la conduite des
EC en termes de recrutement, de rétention et d’adhérence
« J'ai eu des bouffées de chaleur et des sueurs nocturnes cette semaine. Est-ce un effet
secondaire de mon traitement, ou devrais-je m'inquiéter d'autre chose ? »
Bonjour,
Désolé d'apprendre que vous avez des bouffées de chaleur et des sueurs nocturnes. Oui, c'est un effet secondaire très
courant de la thérapie de privation d'androgènes. Il existe des médicaments qui peuvent être utiles pour gérer cela si cela
devient très gênant. Sinon, c'est un effet attendu de ce traitement.
Cher patient,
Les bouffées de chaleur et les sueurs nocturnes que vous ressentez sont probablement des effets secondaires de votre traitement
actuel contre le cancer, en particulier de la thérapie de privation d'androgènes (ADT). L'ADT fonctionne en réduisant les niveaux
d'hormones mâles (androgènes) dans votre corps, ce qui peut aider à ralentir la croissance du cancer de la prostate. Cependant, cette
réduction des androgènes peut également provoquer des symptômes similaires à ceux ressentis par les femmes pendant la
ménopause, tels que des bouffées de chaleur et des sueurs nocturnes.
Il est important de discuter de ces effets secondaires avec votre oncologue lors de votre prochaine visite, car il ou elle pourra
recommander des stratégies ou des médicaments pour aider à gérer ces symptômes. En attendant, vous pouvez essayer certaines
mesures d'auto-soins comme porter des vêtements légers et respirants, utiliser des ventilateurs ou la climatisation, et éviter les
déclencheurs tels que la caféine, l'alcool et les aliments épicés.
N'hésitez pas à nous contacter si vous avez d'autres questions ou préoccupations.
• Les messages de GPT-4 sont acceptables
sans modifications dans 58 % des cas.
Drugs
Madec, J., Bouzillé, G., Riou, C., Van Hille, P., Merour, C., Artigny, M.-L., Delamarre, D.,
Raimbert, V., Lemordant, P., Cuggia, M., 2019. eHOP Clinical Data Warehouse: From a
Prototype to the Creation of an Inter-Regional Clinical Data Centers Network. Stud Health
Technol Inform 264, 1536–1537. [Link]
Accès aux données : Dé-silotter les territoires
Les hubs de données interrégionaux
Accès aux données : Dé-silotter au delà:
Les hubs de données nationaux et européens
Cuggia, M., Polton, D., Weinrib, G., Combes, S., 2018. Rapport Health Cuggia M, Combes S. The French Health Data Hub and the German Medical Informatics Initiatives: Two
Data Hub, mission de préfiguration. Ministère de la santé et des National Projects to Promote Data Sharing in Healthcare. Yearb Med Inform. 2019 Aug;28(1):195-202.
solidarité, Paris. doi: 10.1055/s-0039-1677917. Epub 2019 Aug 16. PMID: 31419832; PMCID: PMC6697511.
Le défi de la qualité
des données
The
GIGO
Effect Garbage in Garbage out
Schwabe, D., Becker, K., Seyferth, M., Klaß, A., Schaeffter, T., 2024.
The METRIC-framework for assessing data quality for trustworthy AI in
medicine: a systematic review. npj Digit. Med. 7, 1–30.
[Link]
Des données qui évoluent
au cours du temps
- Changement de plateforme
- en 2012 : Beckman vers Roche - Changement de
technique en 2023
- Colorimétrie vers
immunonéphélémét
rie
- Sans changement
Valeurs moyennes par mois en g/l
de bornes
Digan, W., Wack, M., Looten, V., Neuraz, A., Burgun, A., Rance, B., 2019.
Evaluating the Impact of Text Duplications on a Corpus of More than 600,000
Clinical Narratives in a French Hospital. Stud Health Technol Inform 264, 103–
107. [Link]
Garantir la sécurité
et la confidentialité
des données
Dé-identification des
Azzouzi, M.E., Coatrieux, G., Bellafqira, R., Delamarre, D., Riou, C., Oubenali,
N., Cabon, S., Cuggia, M., Bouzillé, G., 2024. Automatic de-identification of
French electronic health records: a cost-effective approach exploiting distant
données textuelles
supervision and deep learning models. BMC Medical Informatics and Decision
Making 24, 54. [Link]
Dé-identification
des images médicales
Technique de “defacing”
L'équité et la justice
au cœur des enjeux
de l'IA.
Exemple de l’IA dans la pandémie Covid
Wynants L, Van Calster B, Collins GS, Riley RD, Heinze G, Schuit E, Bonten MMJ, Dahly DL, Damen JAA,
Debray TPA, de Jong VMT, De Vos M, Dhiman P, Haller MC, Harhay MO, Henckaerts L, Heus P, Kammer M,
Kreuzberger N, Lohmann A, Luijken K, Ma J, Martin GP, McLernon DJ, Andaur Navarro CL, Reitsma JB,
Sergeant JC, Shi C, Skoetz N, Smits LJM, Snell KIE, Sperrin M, Spijker R, Steyerberg EW, Takada T, Tzoulaki I,
van Kuijk SMJ, van Bussel B, van der Horst ICC, van Royen FS, Verbakel JY, Wallisch C, Wilkinson J, Wolff R,
Hooft L, Moons KGM, van Smeden M. Prediction models for diagnosis and prognosis of covid-19: systematic
review and critical appraisal. BMJ. 2020 Apr 7;369:m1328. doi: 10.1136/bmj.m1328. Update in: BMJ. 2021
Feb 3;372:n236. doi: 10.1136/bmj.n236. Erratum in: BMJ. 2020 Jun 3;369:m2204. doi: 10.1136/bmj.m2204.
PMID: 32265220; PMCID: PMC7222643.
• Algorithmes
• Data driven
• Permette l’aide à la
décision individuelle
• Combinent et
définissent le poids de
différentes variables
d’entrées comme
l'âge, le signe et
symptôme, la biologie,
l’imagerie
Des modèles performants mais pour la
plupart biaisés
C-Statistique
Des biais discriminatoires …
§ Nécessité de comprendre le
raisonnement des modèles d'IA pour
les rendre plus acceptables et fiables.
§ Méthodes d'explicabilité : cartes
d'importance des variables, exemples
contrefactuels, techniques de
visualisation.
§ Enjeu de la confiance et de la
transparence vis-à-vis des patients.
Exemples de méthode d’explicabilité en anatomopathologie
Cabitza, F., Zeitoun, J.-D., 2019. The proof of the pudding: in praise of a
culture of real-world validation for medical artificial intelligence. Annals of
Translational Medicine 7, 161–161. [Link]
Impact de l’IA sur le
• Agnès Buzyn (ancienne ministre de la Santé) :
"L’intelligence artificielle peut être un allié puissant pour compenser le manque de médecins,
notamment dans les déserts médicaux, en améliorant le diagnostic à distance et en optimisant la
Système de santé gestion des ressources de santé. Mais elle ne remplacera jamais la présence humaine
indispensable dans le soin."
Vers une délégation de tâche via l’IA
des tchatbots
Empathiques
Ayers, J.W., Poliak, A., Dredze, M., Leas, E.C., Zhu, Z., Kelley, J.B.,
Faix, D.J., Goodman, A.M., Longhurst, C.A., Hogarth, M., Smith,
D.M., 2023. Comparing Physician and Artificial Intelligence
Chatbot Responses to Patient Questions Posted to a Public
Social Media Forum. JAMA Intern Med 183, 589–596.
[Link]
Coopérants
Google propose C oDoC, une IA médicale qui sait
quand s'en remettre à des experts humains.
• Travail multidisciplinaire
indispensable
Conclusions
• L’IA en santé est un secteur
passionnant et un domaine en
pleine expansion
• Nombreuses perspectives
professionnelles riches et très
variés
Kinoshita, Fumihiko, Tomoyoshi Takenaka, Takanori Yamashita, Koutarou Matsumoto, Yuka Oku, Yuki Ono, Sho
Wakasu, et al. « Development of Artificial Intelligence Prognostic Model for Surgically Resected Non-Small Cell Lung
Cancer ». Scientific Reports 13, no 1 (21 septembre 2023): 15683. [Link]