Causal Foundation Models for Multimodal Precision Treatment-Response Discovery

Project Description

Precision medicine increasingly relies on heterogeneous data spanning electronic health records (EHRs), multi-omics, and longitudinal functional measurements from wearable and mobile sensors. Yet current multimodal foundation models primarily learn associations for prediction rather than uncovering how treatments causally affect outcomes across heterogeneous patients. This project will develop a causal multimodal foundation model for treatment-response discovery, integrating EHR trajectories, molecular profiles, and high-frequency functional data within a unified representation and causal inference framework.

The research will address three fundamental challenges: (1) learning transferable patient representations across modalities with different temporal scales, missingness patterns, and measurement processes; (2) distinguishing treatment effects from correlations arising from confounding and clinical selection; and (3) discovering patient-specific and subgroup-specific treatment-response heterogeneity. The fellow will combine multimodal representation learning and foundation models with causal inference, longitudinal modeling, and operations research methods for treatment decision-making.

The resulting framework will enable researchers to move beyond predicting what will happen toward estimating what would happen under alternative treatments, providing a methodological foundation for individualized treatment-response discovery and evidence-based precision medicine.

Start Date

September 1, 2027

Postdoc Qualifications

We seek candidates with a PhD in operations research, industrial engineering, biomedical/health informatics, computer science, statistics, biostatistics, computational biology, biomedical engineering, or a closely related field. Strong quantitative and computational skills are essential. Particularly relevant expertise includes one or more of the following: causal inference, machine learning or foundation models, multimodal learning, longitudinal/time-series modeling, optimization and sequential decision-making, electronic health record analytics, computational genomics/multi-omics, or digital health/wearable data.

The ideal candidate will have demonstrated ability to develop new quantitative methodology, work with large-scale biomedical datasets, and collaborate across engineering, data science, and health sciences. Strong programming skills in Python and experience with modern machine-learning frameworks are desirable.

Co-advisors

Nan Kong, Industrial Engineering and Biomedical Engineering, nkong@purdue.edu
Serena Jingchuan Guo, Pharmacy, serena.guo@purdue.edu

Bibliography

A Foundation Model for Wearable Movement Data in Mental Health Research.
Ruan FY, Zhang A, Oh JY, Jin S, Jacobson NC. IEEE J Biomed Health Inform. 2026 May 19;PP. doi: 10.1109/JBHI.2026.3694809. Online ahead of print.

Tang H, Donahoo WT, DeKosky ST, Lee YA, Kotecha P, Svensson M, Bian J, Guo J. Heterogeneous treatment effects of GLP-1RAs and SGLT2is on risk of Alzheimer's disease and related dementia in patients with type 2 diabetes: Insights from a real-world target trial emulation. Alzheimers Dement. 2025 Jun;21(6):e70313. doi: 10.1002/alz.70313. PubMed PMID: 40448382; PubMed Central PMCID: PMC12125485.

Prosperi M, Rancati S, Guo Y, Guo J, Radwan RM, Salemi M, Bian J, Bhasuran B, Egelund EF, Janelle J, Manavalan P, Maranchick N, Marini S, He Z. Consistency in causal reasoning for large language models in scenarios of HIV antiretroviral treatment, drug interactions, and side effects. NPJ Digit Med. 2026 May 27;. doi: 10.1038/s41746-026-02771-7. [Epub ahead of print] PubMed PMID: 42204276.

Mu Du, Hongtao Yu, Nan Kong, Transfer reinforcement learning for mixed observability Markov decision processes with time-varying interval-valued parameters and its application in pandemic control. INFORMS Journal on Computing. Vol. 37, No. 2, pp. 315 - 337.

Haiyan Lu, C.-C. Yang, Nan Kong, Semi-parametric Imputation for Stratified Treatment Choice under Partially Scarce Response Data. Submitted to Production and Operations Management.