AI for the Proper Use of Medications
summary
Vidal built a Data Science platform to semi‑automatically extract drug indications, contraindications and adverse effects from about 10,000 RCP documents using 19 years of pharmacist annotations, producing roughly 10,000 deep models (≈20k including similarity models) and reaching 85–90% F1. Extraction uses CamemBERT with convolutional heads for transfer learning, a Ratcliff‑Obershelp string similarity method and rule‑based postprocessing; pharmacists validate monthly suggestions through a UI and CI/CD plus retraining are orchestrated with Airflow, Jenkins, S3, DVC and Kubernetes with continuous feedback into the knowledge base.
AI-written summary · about
chapters (23)
- 0:00Introduction
- 0:49THE MEDICINE IS NOT A PRODUCT LIKE ANY OTHER
- 2:09Assistance for proper medication use
- 3:38Exploiting knowledge
- 5:52Concept extraction
- 7:47Automatic concept extraction
- 8:07The project
- 9:24The feature team
- 9:49Learning database
- 10:32Extraction algorithms
- 11:11Similarity-based approach
- 15:56BERT: Transfer Learning
- 16:58Embeddings
- 21:21Training dataset
- 23:40Convolution
- 25:44Common denominator
- 26:57K Class CommonDenominator Model
- 27:40Business rules method
- 28:38Hybrid method: An RCP annotation
- 29:30How many models?
- 30:08Evaluation
- 30:39The Data Science platform with the pipeline and feedback loop
- 44:09DEVOX France
related talks
28mAI in healthcare - Deep learning for precision psychiatry
47mAI for making decisions in clinical medicine
17mAI Innovation in Health
35mCan AI solve the biggest problem in drug discovery? Time!
29mAI for COVID-19: An online virtual care approach
44mRAG for a Medical Company: Technical and Product Challenges AI & Data
30mAI in Drug Discovery
25mAI for healthcare: Scaling Access and Quality of Care for Everyone