Himadri S. Pandey · PhD Candidate, Machine Learning, Georgia Institute of Technology

Research

Who should get help, and what kind of help, when there isn't enough to go around?

I study two linked decisions: how to allocate scarce clinical attention, and which information is worth collecting before allocating it. I combine machine learning, optimization, and sequential decision-making, and I build these models with the clinicians who will use them.

Allocating care · Choosing what to learn · Where this leads

Allocating care

Hospital wards can respond to a deteriorating child with closer monitoring, a specialist ICU team, or early ICU transfer, and each response has its own limited capacity. My work with Children’s Healthcare of Atlanta asks who should receive which response.

Who gets watched?

Allocating Scarce Escalation Resources on Pediatric Wards with Neural Index Policies

Job market paper · Under review, M&SOM Request manuscript

In a retrospective evaluation on historical CHOA data, the policy recommends an escalation before 87.4% of critical or emergent ICU transfers (vs. 79.5% for the stronger baseline), while using 13.5% less monitoring capacity, with a median lead time of 19.2 hours.

Study details

I model the ward as a multi-action restless bandit and develop a neural index policy that learns priorities across patients and response types, then assigns responses within each capacity limit. The results above come from held-out historical records; we are planning a silent deployment with CHOA to evaluate the policy prospectively. A companion paper I coauthored, on early prediction of pediatric ICU transfer, has been submitted to INFORMS Journal on Data Science.

Choosing what to learn

Good allocation depends on what clinicians know, and collecting that information takes staff time. I study this in concussion diagnosis with the NCAA–DoD CARE Consortium.

Which tests fit in the time available?

So Many Tests, So Little Time: Operationally Constrained Machine Learning for Diagnostic Assessment under Time Constraints

IISE Transactions, 2026 Paper

Designs test batteries that take more than 97% less time to administer than batteries used in practice, while keeping AUC above 0.93.

Study details

A chance-constrained mixed-integer program that selects tests and fits an interpretable diagnostic model together, treating each test’s administration time as uncertain. Clinicians set both the time budget and the acceptable risk of exceeding it. Winner, INFORMS MIF Student Poster Competition (2024).

Who needs a baseline?

Quantifying the Diagnostic Utility of Baseline Testing in Concussion Management

The American Journal of Sports Medicine, 2025 Paper

In our CARE Consortium sample, adding pre-season baseline scores did not change the classifier’s diagnosis for 86.7% of concussed athletes.

Study details

Symptoms fluctuate day to day, so a baseline can add noise as well as information. By comparing classifiers built with and without baselines, I measure when a baseline changes the diagnosis. A follow-up study finds routine baseline testing costs nearly 600 staff hours per 1,000 athletes each season.

Ongoing work. Optimal logistic regression trees for cost-aware diagnosis jointly chooses how to group patients, which tests each group receives, and how each group’s results determine a diagnosis. In preliminary concussion results, it identifies athletes with chronic medical conditions as a distinct group with its own diagnostic model.

Where this leads

Care when clinicians' responses evolve. Recommendations change how clinicians respond to later ones. I will design allocation policies that account for this shared, evolving response.
What to learn before allocating care. Deciding whom to assess next and when to act, when assessment and care draw on the same scarce staff.
Outreach and continuing mental health care. Working with outreach organizations, mental health providers, and people with lived experience of homelessness on how to balance new contacts with follow-up when sustained care depends on capacity across services. This builds on my mental health leadership at Georgia Tech.