Outpatient care is central to healthcare systems, with general practitioners and specialists Research focus.

The extensive real-world data captured by private practitioners — consultation reports, imaging, referral letters, prescriptions, and lab results — offers a unique opportunity to explore how large language models (LLMs) and generative approaches can enhance diagnostic processes while incorporating elements of clinical reasoning.

This thesis aims to develop a multimodal, text-centric, and longitudinal generative model to predict optimal next actions in outpatient care pathways. The goal is to support diagnostic decisions by recommending actions (ordering tests, specialist referrals) that lead to correct diagnoses through optimal pathways, while providing justification through clinical reasoning elements.

The research will utilize timestamped sequences of both structured data (such as lab results, prescriptions) and unstructured clinical texts (consultation summaries, referral letters), adapting the "episode of care" framework to outpatient diagnosis contexts.

Scientific Challenges include Dataset Curation and Evaluation, Multimodal Modeling, Uncertainty Calibration, Clinical Reasoning and Guideline-Compliant Planning, Integration within Agentic Systems.

What you will do

Establish a comprehensive review of literature.
Choose and evaluate relevant baselines, deliver a documented benchmark.
Publish at least two scientific contributions, either in a scientific international journal or a conference. Example of targeted venues: journal about application of AI in medicine such as NPJ Digital Medicine, the Lancet Digital Health, AIIM, JAMA Open Network, JBI, JAMIA; for AI conferences:, NeurIPS, AAAI, ICML, ICLR, IJCAI/ECAI.
Perform evaluation on both synthetic and real world datasets (Doctolib).
Implement a prototype that can be easily transferred to Doctolib’s suite of products.

Who you are

You have

Technical proficiency in Python, and with basic data science libraries (pandas, scikit-learn, etc.)
Experience in ML/DL (Pytorch, JAX, …) and with programmatic use of LLMs
Scientific background
Excellent communication skills in particular in scientific writing
Sincere motivation for the healthcare domain.

Supervision

Academic co-supervisors: Adrien Coulet, PhD, INRIA Researcher HDR, Ivan Lerner, MD PhD, PHU Université Paris Cité, AP-HP
Research unit: Inria, Inserm, Université Paris Cité, U1346 HeKA

Industrial supervisors: Fajwel Fogel, PhD, Data Science Manager; Nicolas Barascud, PhD, Principal Data Scientist
Company: Doctolib

General information

The HEKA team at Inria conducts multidisciplinary research at the intersection of artificial intelligence and healthcare.

Doctolib is France's leading digital health platform, connecting over 400,000 healthcare professionals with approximately 80 million patients across Europe. The company provides software solutions that streamline healthcare administration, appointment scheduling, and telehealth services. As a health technology pioneer, Doctolib captures extensive real-world healthcare data while maintaining strict privacy standards. Beyond its core booking services, the company is increasingly investing in artificial intelligence and data science to develop innovative tools that support clinical decision-making and improve healthcare delivery, positioning itself at the forefront of digital transformation in European healthcare.

The PhD candidate will share their time (half-half) between Doctolib headquarters (54 quai Charles Pasqua, 92300 Levallois-Peret) and the HeKA team offices in PariSanté Campus (2-10 rue d’Oradour-sur-Glane, 75015 Paris).