11/04/2025
By Danielle Fretwell

The Francis College of Engineering, Department of Mechanical Engineering, invites you to attend a Doctoral Dissertation defense by Elyas Irankhah on: "Machine Learning and Causal Inference for Personalized Outcome Modeling in Traumatic Brain Injury."

Candidate Name: Elyas Irankhah
Degree: Doctoral
Defense Date: Wednesday, November 12, 2025
Time: 10 a.m.-noon
Location: Southwick 240

Committee:

  • Advisor: Kelilah L. Wolkowicz, Ph.D., Assistant Professor, Department of Mechanical & Industrial Engineering, UMass Lowell
  • Co-Advisor: Mohammad Arif Ul Alam, Ph.D., Assistant Professor, Miner School of Computer and Information Sciences, UMass Lowell
  • Alessandro Sabato, Ph.D., Associate Professor, Department of Mechanical & Industrial Engineering, UMass Lowell
  • David Claudio, Ph.D., Associate Professor, Department of Mechanical & Industrial Engineering, UMass Lowell
  • Sashank Narain, Ph.D., Assistant Professor, Miner School of Computer and Information Sciences, UMass Lowell

Abstract:
Traumatic Brain Injury (TBI) affects over 69 million people annually and remains one of the most challenging conditions in clinical care. Despite extensive research, treatment decisions remain limited by three major issues: heterogeneous recovery, confounding in observational data, and lack of transparent (validated AI models). This dissertation addresses these gaps through a three phase framework that integrates machine learning, causal inference, and large language models (LLMs) to improve individualized outcome modeling and rehabilitation strategies.

Phase 1 – Cognitive Assessment and Detection
Supervised learning models were applied to the Virtual Reality-Cognitive Assessment Tool (VR-CAT) dataset to evaluate whether task based performance metrics could distinguish TBI from control groups. The models achieved strong accuracy and clear interpretability. Task-based features such as accuracy and response time supported early, non-invasive TBI detection.

Phase 2 – Causal Inference for Interventions
Using data from the Traumatic Brain Injury Model Systems (TBIMS), a unified causal framework combined Outcome-Adaptive LASSO, CGNN, IPW, and CEVAE to estimate treatment effects. Analyses revealed that early rehabilitation and longer rehabilitation duration significantly improved productivity, social outcomes, and life satisfaction. Craniotomy showed modest short-term functional decline. Targeted recovery support is therefore recommended.

Phase 3 – LLM-Enhanced Causal Validation
A large-scale study using 24,939 patients from the All of Us Research Program incorporated medical embeddings and counterfactual validation through LLMs. Findings confirmed that rehabilitation within three days reduced 30-day readmissions by about 20%, with results consistent across causal and generative AI methods.

Significance
This dissertation introduces a practical pathway toward precision neurorehabilitation and combines advanced AI methods with clinical insight to improve recovery prediction and treatment planning after TBI.