Himadri S. Pandey · PhD Candidate, Machine Learning, Georgia Institute of Technology
Publications
Job market paper · Selected talks · Google Scholar
Job Market Paper
Who Gets Watched? Allocating Scarce Escalation Resources on Pediatric Wards with Neural Index Policies
Finalist, Lee B. Lusted Student Prize Competition (QMTD), SMDM 2026
Abstract
Problem definition: Hospitals use separate pathways to identify and respond to deteriorating ward patients. At Children’s Healthcare of Atlanta (CHOA), these include Watcher surveillance, Critical Care Outreach Team assessment, and transfer to the intensive care unit. When several patients need attention simultaneously, however, no common process determines who should receive which response under limited staffing and capacity. Methodology/results: We model ward escalation decisions as a multi-action restless multi-armed bandit and develop the Multi-action Index Learning and Optimization Policy (MILOP), which trains a neural network to generate a high-quality index policy using decision-focused learning. Using historical CHOA data, MILOP recommends an escalation before 87.4% of critical or emergent ICU transfers, compared with 79.5% for the stronger baseline, while using 13.5% less monitoring resources, and provides a median of 19.2 hours of lead time. Managerial implications: Hospitals may improve escalation decisions by coordinating existing pathways rather than managing each separately or using risk scores as stand-alone decision rules.
Journal Articles
Winner, MIF Student Poster Competition, INFORMS 2024
Abstract
Practitioners must often make diagnostic decisions under operational constraints; for example, clinicians diagnosing concussion may assess patients under time limits imposed by athletic, emergency department, or military settings. We generate interpretable ML models that enforce such constraints, modeling battery design as a chance-constrained mixed integer program with reformulations for tractability. On data from the NCAA–DoD CARE Consortium, the resulting batteries reduce administration time by more than 97% relative to batteries used in practice while maintaining AUC above 0.93.
Under Review and Submitted
Early Prediction of Pediatric ICU Transfer: A Comparative Study of Tabular and Time Series Machine Learning Models
Working Papers
Rank, Then Allocate: A Sequential Knapsack Framework for Feature-Driven Personalization
Who Gets Which Test? Optimal Logistic Regression Trees for Cost-Aware Diagnosis
Abstract
Jointly chooses how to group patients, which tests each group receives, and how each group’s results determine a diagnosis, solved exactly by dynamic programming. Preliminary concussion results identify athletes with chronic medical conditions as a distinct group with its own diagnostic model.
Accuracy–Calibration Tradeoffs in an Operationally Constrained Machine Learning Model for Concussion Risk Prediction Among Service Academy Cadets
Book Chapters
Conference Proceedings
Selected Talks and Presentations
-
Nov 2026
“Neural Index Policies for Multi-Action Restless Multi-Armed Bandits with Heterogeneous Budgets”
-
2026
“Resource-Aware Decision Making for Pediatric Post-Surgical Monitoring via Neural Index Policies”
-
2026
“So Many Tests, So Little Time: Operationally Constrained Machine Learning for Diagnostic Assessment under Time Constraints”
-
2024
“Time-Constrained Concussion Diagnosis: An Analysis of the NCAA-DoD CARE Consortium Dataset”
-
2024
“Value of Baseline Testing in Concussion Diagnosis”
The full list of presentations is in the CV.