Fair AI
BMDS 223 — Deploying and Evaluating Fair AI in Healthcare
One lecture and one hands-on coding workshop on auditing language models and clinical data for bias.
About
I am a postdoctoral scholar working with Prof. Tina Hernandez-Boussard in the Division of Computational Medicine, Department of Medicine, Stanford University. My research is on responsible reasoning for AI-assisted clinical decision making, at the intersection of biomedical AI, machine learning and natural language processing.
I design and evaluate clinical AI systems on real-world patient data — structured EHR and clinical notes — across Parkinson’s disease (L3S Research Center & Hannover Medical School), oncology (Wipro GE Healthcare) and postoperative pain management (Stanford Medicine). My work spans reliability evaluation of medical LLMs, RAG systems and agents; guideline-grounded clinical decision support; efficient pretraining of biological foundation models; and domain adaptation for medical text.
I received my PhD in Computer Science and Engineering from IIT Kharagpur (thesis: Domain Adaptation for Medical Language Understanding). My work has appeared at venues including ACL, IJCAI, SIGIR, EMNLP, ECAI, LREC, CIKM and Frontiers in Artificial Intelligence.
Education
-
IIT
KGP2025PhD, Computer Science and Engineering
Indian Institute of Technology Kharagpur · Thesis: Domain Adaptation for Medical Language Understanding
-
IIT
KGP2019MS (by Research), Computer Science and Engineering
Indian Institute of Technology Kharagpur · Thesis: Computational Approaches for Online Monitoring · GPA 9.21
-
KGEC
2017
B.Tech, Computer Science and Engineering
Kalyani Government Engineering College · Thesis: Understanding Email Interactivity and Predicting User Response to Email · GPA 8.61
Research
My central question is how prescriptive knowledge sources — above all, clinical guidelines — can be used to build scalable, data-driven approaches that ensure the evidence quality and reliability of AI-assisted clinical decision making. I work on this in two connected directions.
Reliability measurement. I study how model reasoning varies under controlled, clinically grounded counterfactuals, and use that variation to estimate how a system will behave after deployment. This includes measuring invariance and sensitivity in synthetic patient trajectories for evaluating patient-to-trial matching (TrialEval), counterfactuals built on clinical risk factors (Q-Pain++), and trajectory-level metrics for rare, compounding failures in sequential clinical agents (cascading tail risk).
Reliability mitigation. Grounding model context in the relevant guideline details leads to more accurate dosage recommendations and better robustness to decision-space framing. In parallel, I build models that compute guideline deviations in real-world patient care trajectories from Stanford Health Care and MIMIC data, to ask whether a deviation leads to better outcomes or avoidable harm.
Research questions I’m interested in
- How do real-world patient trajectories deviate from clinical guidelines — and when does a deviation help or harm?Can we turn guidelines into computable pathways and measure deviations at scale over longitudinal EHR data and clinical notes?
- How can we estimate a clinical LLM’s post-deployment behaviour before it is deployed?Which controlled, clinically grounded counterfactuals reveal whether a model’s reasoning is invariant where it should be and sensitive where it must be?
- Why are binary safety audits inadequate for frontier clinical models?How much of a model’s treatment decision is driven by how the decision space is framed, rather than by the patient in front of it?
- How do small errors compound in sequential clinical agents?Can trajectory-level tail-risk metrics surface rare, escalating harms that per-step accuracy hides — for example in opioid prescribing after surgery?
- Does uncertainty quantification flag the errors that matter clinically?How do we build decision-aware evaluations and scale-independent uncertainty measures that prioritise high-risk mistakes?
- How can we make foundation models for medicine and biology cheaper to adapt?Can vocabulary adaptation and smarter masking cut pretraining and fine-tuning cost without sacrificing faithfulness on specialised text and genomic sequences?
Methods, frameworks & datasets
FrameworkQ-Pain++
Factorial stress-testing and multi-modality evaluation of clinical LLMs. An audit of five models exposed a decision-space framing effect driving ~92% of treatment-decision changes; guideline grounding improves robustness.
PreprintUnder review
FrameworkCascading Tail Risk
Trajectory-level safety evaluation for sequential clinical LLM agents, with Tail Harm Escalation and Tail-Restricted Wasserstein Distance metrics, demonstrated on postoperative opioid prescribing.
Under review
MethodGeneMask
NPMI-guided masking for masked language modelling of DNA sequences. 10× faster pretraining for few-shot gene sequence classification with DNABert and LOGO.
MethodCurriculum GeneMask↳ from GeneMask
Adaptive curriculum masking for DNA transformers, improving genome-understanding benchmarks in both full-data and few-shot settings.
MethodMEDVOC
Vocabulary adaptation for medical summarization, with fragment-score-based hyperparameter search that cuts fine-tuning time from 450 days to about 45 hours.
MethodVocabReplace-Then-Expand
Parameter-efficient vocabulary adaptation for LLMs: adds domain tokens while replacing under-trained ones. 35–55% less training time than continual pretraining, 37% fewer parameters.
MethodAdaptBPE
Modifies BPE initialisation so domain-specific terminology is preserved during fine-tuning of pretrained language models.
MethodDecision-tree PD subtyping
Interpretable Parkinson’s disease subtyping on the MJFF LRRK2 cohort (1,500+ patients); six subtypes externally validated on the MDS dataset.
DatasetG4UE — GPT-4 USMLE errors
GPT-4 responses to USMLE questions with a medical-expert error taxonomy of 8 failure modes and span-level annotations by 44 medical experts.
DatasetLongTailQA
A disambiguated long-tail QA benchmark addressing 8–30% entity ambiguity in existing datasets; query disambiguation gives RAG models up to 24.7% accuracy gains.
DatasetPhysician intent in dialogues
Annotated doctor–patient dialogues for physician intent modelling; intent filtering improves dialogue summarization.
DatasetMulti-disease trial search set↳ from MPSS
~25 consumer queries with ~95 annotated clinical trials each, released with a metapath-based trial search model over a 750K-node PubMed–trial graph.
Publications
Under review
-
Tail
riskCascading Tail Risk: Trajectory-Level Safety Evaluation for Sequential Clinical LLM Agents
Under review
-
Q-Pain
++Beyond Binary Audits: Does Guideline Grounding Mitigate Decision-Space Framing in LLM Opioid Prescribing?
Introduces Q-Pain++, a factorial stress-testing framework auditing five clinical LLMs.
Under review
-
UQ
A decision-aware evaluation suite for uncertainty quantification in clinical ML
Proposes the scale-independent Uncertainty Index; shows standard UQ methods fail to flag high-risk errors across three real-world datasets.
Under review
2026
-
ACL
2026 -
LREC
2026
2025
-
Front.
AI
-
ECAI
2025 -
ACL
2025 -
WSDM
2025 -
WSDM
DayA Systematic Evaluation of Single-Cell Foundation Models on Cell-Type Classification Task
WSDM Day 2025
2024
-
SIGIR
2024Beyond Accuracy: Investigating Error Types in GPT-4 Responses to USMLE Questions
SIGIR 2024 · Resource & Reproducibility track
-
IJCAI
2024 -
EMNLP
2024 -
ECAI
2024
2023
-
ECAI
2023
-
ICDH
2023
Earlier work (2021–2018) →
2021
-
CIKM
2021
-
TWEB
2021
2019–2018
-
WebSci
2019
-
CSoNet
2019 -
CICBA
2018
Experience
-
SU
Postdoctoral Scholar, Division of Computational Medicine
Stanford University · Stanford, CA · 09/2025–present
- Lead three projects identifying failure modes in clinical LLMs — decision-space framing, cascading tail risk, and robustness under controlled clinical counterfactuals — and mitigations grounded in clinical guidelines.
- Build and evaluate LLM-based clinical decision support for guideline adherence over real-world patient trajectories and for synthetic cohort generation to evaluate patient-to-trial matching systems.
- Develop computable guideline pathways and deviation analysis over Stanford Health Care and MIMIC data (structured EHR and clinical notes).
-
GE
PhD AI Intern, HealthCare Technology and Innovation Center
Wipro GE Healthcare · Bangalore, India · 11/2024–05/2025
- Built a generative-AI dashboard over oncology patient data (Chroma, LangChain) for semantic search across complex patient histories.
- Developed representation learning for medical-imaging-equipment sensor logs for anomaly detection and predictive maintenance; one patent filed.
-
L3S
Research Associate
L3S Research Center, Leibniz University Hannover · Germany · 01/2021–06/2023
- Led a collaboration with Hannover Medical School on Parkinson’s disease subtyping using the MJFF LRRK2 dataset; six novel subtypes externally validated on the MDS dataset.
- Developed GeneMask, PMI-based token masking for DNA transformers with a 10× speedup for few-shot gene sequence classification.
-
ADBE
Research Intern, Big Data Experience Lab
Adobe Systems India · Bangalore, India · 05/2018–06/2018
- Built brand-personality detection and brand-aligned sentence recommendation models; led a quality-controlled Amazon Mechanical Turk annotation study. Resulted in ACM WebSci 2019 and ACM TWEB 2021 papers.
Applied & translational work
- 2026
MIT Hacking Medicine — Boston GrandHackCo-developed AuriCare, a holistic pain-management decision-support concept. DemoBlogEvent
- 2025–
Computable guideline pathways at StanfordComputing deviations in patient care trajectories with Stanford Health Care and MIMIC data.
- 2025
Patent filing — Wipro GE HealthcareCo-inventor: deep-learning representation learning for medical-imaging-equipment sensor logs, enabling anomaly detection and predictive maintenance.
- 2024
Funded translational project — AI4ICPS, IIT KharagpurMy PhD work on interpretable clinical trial search continues as a funded project led by a co-author. ProjectPoster
Teaching
Lecture and workshop material is linked where available.
Stanford University
Leibniz University Hannover · Germany
NLP
Natural Language Processing
Retrieval
Foundations of Information Retrieval
IIT Kharagpur
Programming
Teaching Assistant
- Programming and Data Structure Lab · Spring 2018
- Information Retrieval · Autumn 2020
- Computing Lab 1 · Autumn 2023
- Department Website Team · Autumn 2018–Spring 2020
Tutorials & invited talks
- Jan 2026
Foundation Models for Medical TextAI Foundation Models in Biomedicine course, Leibniz University Hannover Slides
- Jan 2026
Building a Career in AI for Industry and AcademiaGoogle Developer Group on Campus, Kalyani Government Engineering College Slides
- Dec 2025
Vocabulary Adaptation Strategies for Finetuning Medical Language ModelsBreakfast Talk series, Microsoft Research India, Bangalore Slides
- Aug 2025
Improving Training Efficiency of DNA Foundation Models1st Symposium on Intersections: Statistical Genomics and Complex/Chronic Disease Biology, BRIC–NIBMG, Kalyani
- Mar 2025
Tutorial — Building Trustworthy AI Models for MedicineWSDM 2025 (3-hour tutorial) WebsiteSlides
- Dec 2024
Foundation Models for Medical TextAI Foundation Models in Biomedicine course, Leibniz University Hannover Slides
Mentoring & service
Mentoring
- 3 MS students (2026)Bioengineering and Biomedical Data Science at Stanford; Biostatistics at Yale.
- 1 PhD thesis, IIT Kharagpur (2026)Efficient vocabulary adaptation of general-purpose language models to the medical domain.
- 3 M.Tech theses, IIT KharagpurClinical trial search (2019), query-focused medical summarization (2020), medical knowledge graphs for NLP (2021).
- 1 B.Tech thesis, IIT Kharagpur (2022)Question generation for QA-based summary evaluation.
Reviewing & memberships
- CORE A* conferencesNeurIPS 2026, FAccT 2026, AAAI 2023–2026, ACL Rolling Review (2025–2026 cycles), ACL 2023, EMNLP 2022–2023.
- CORE A conferencesCOLING 2025, ML4H 2024, EACL 2023, ECML-PKDD 2020.
- Medical AI & informaticsAMIA 2026, JAMIA, Frontiers in AI, Frontiers in Genetics, KAIS, IEEE TCDS, Artificial Intelligence in Medicine.
- MembershipsSigma Xi, Association for Computational Linguistics, AMIA.