Clinical Applications of Causal Inference, Machine Learning, and Pharmacovigilance Safety Guardrails
Name
hakim-joehakim-phd-hst-2026-thesis.pdf
Size
4.82 MB
Format
Adobe PDF
Checksum (MD5)
6cb7d389e4ac3f7f6d45a2f8c6bf2262
Author(s)
Hakim, Joe Bahij
Advisor(s)
Beam, Andrew
Date Issued
May 2026
Publisher
Massachusetts Institute of Technology
Abstract
In this thesis, I explore and evaluate non-purpose-collected medical data (claims, safety reports, longitudinal ICU records) and make use of these data in both evaluating specific evidential claims and prototyping systems that make their usage help guide policy and safety decisions of different medical interventions.
There are three primary works in this thesis. In the first aim, we look at 17-hydroxy progesterone caproate, and do a target trial using a national insurance claims database to assess the effectiveness in it's specific indication; in 4,422 eligible pregnancies (568 treated), the matched estimate showed no beneficial effect (RR 1.1, 95\% CI 0.9-1.4), and this null signal remained stable across subgroup and sensitivity analyses.
In the second aim, we constructed a system of guardrails and an eavluation framework for the usage of LLMs in performing a subtask within pharmacogicilance (PV). Our fine-tuned LLM, as assessed by human reviewers, had produced ``never event'' like errors in processing intake case safety reports, and as such we proof-of-concept probabilistic and deterministic guardrails to help robustify and double check outputs of LLMs doing safety-critical PV work.
In the third aim, we report preliminary work in new clustering techniques with applications to causal inference for longitudinal data parametrized by many clinical covariates. We employ effective sample size screening to suites of dynamically defined policies and introduce a tunably clinically conformant feature clustering, balancing data and user-specified feature group representations, and demonstrate this in a sepsis ICU longitduinal cohort.
In sum, this thesis is a suite of results that variably makes use of real world data with causal machine learning and large scale safety-crtiical data processing in order to make bounded, auditable claims about intervention efficacy and safety in settings where the data are inherently messy (and the decisions are high-stakes).
MIT Department
Harvard-MIT Program in Health Sciences and Technology
Terms of Use
In Copyright - Educational Use Permitted
Copyright retained by author(s)
Persistent DSpace Link