Sayantan Kumar

Research profile

Research summary

Sayantan Kumar is a researcher at the United States National Library of Medicine working at the intersection of clinical natural language processing, biomedical informatics, and machine learning in healthcare. The OpenAlex-derived snapshot records an h-index of 8, an i10-index of 7, and 42 works in the most recent three-year window, with annual output accelerating sharply: 8 works in 2024, 15 in 2025, and 19 in 2026, which drives the maximum activity score of 100. Topical analysis concentrates on machine learning in healthcare (28 works), dementia and cognitive impairment research (18), and topic modeling (18), with further activity in Alzheimer's disease and multimodal machine learning. The Evolve-reviewed catalog comprises three 2026 arXiv preprints, all centered on clinical timelines and temporal reasoning: GAVEL, a case-report-grounded LLM judge for comparing extracted timelines against their source reports; a paired benchmark showing that future-data exposure induces hindsight bias in LLM clinical judgment; and a first-author study on UID-preserving multimodal timeline reconstruction with source-grounded adjudication. Together they trace a coherent program: making LLM-based extraction and evaluation of clinical time information verifiable against source evidence. The reproducibility dimension scores 55 (mean across three medium-readiness papers), collaboration scores 39 on a compact six-coauthor network, and focus scores 0 because the field distribution splits evenly across three closely related clinical NLP labels. Citation-based output and impact metrics are not yet available and remain unscored.

Research keywords

Highlights

• GAVEL: grounding LLM judgment of clinical timelines in case reports

  The 2026 GAVEL framework compares extracted clinical timelines against their source case reports using a grounded LLM judge, with validation built on a Bradley-Terry comparison setup. Evolve rates the validation as well executed but cautions that merge gains reuse the judging instrument itself and that judge recall remains unmeasured.

• Paired benchmark exposes hindsight bias in LLM clinical temporal reasoning

  This 2026 paired benchmark shows that exposing the full clinical timeline shifts large language model answers toward hindsight traps without improving accuracy. Evolve acknowledges the paired design but notes that GPT-generated labels and LLM-based judging temper the strength of the evidence.

• First-author work on UID-preserving multimodal timeline reconstruction

  Kumar's first-author 2026 study reconstructs clinical timelines while preserving unique identifiers, combining multimodal EHR signals, discharge summaries, retrieval-augmented generation, and source-grounded adjudication with the GAVEL judge. Evolve credits rigorous ablations but flags the small n=40 sample and a 40%-versus-43% recovery-claim mismatch.

Timeline

1. 2021Early sparse record: OpenAlex logs 3 works in 2021, following a single entry in 2020, in the period before the recent output acceleration.
2. 2024Output accelerates to 8 recorded works, beginning the steep climb that ultimately tops out the activity dimension at 100.
3. 202515 recorded works in the year, nearly doubling the prior year's count within a 42-work recent three-year total.
4. 202619 recorded works so far, including all three Evolve-cataloged arXiv preprints on clinical timelines: GAVEL, the hindsight-bias benchmark, and the first-author UID-preserving reconstruction study.

Papers in the Evolve catalog

• 2026-09-11

  Grounded Adjudication of Variations across Extracted TimeLines (GAVEL): Comparing Clinical Timelines Against Their Case Reports

• 2026-09-11

  Hindsight Bias in Clinical Temporal Reasoning: How Future Data Exposure Affects Large Language Model Judgment

• 2026-09-11first author

  Anchoring Clinical Events in Time: UID-Preserving Multimodal Reconstruction and Source-Grounded Adjudication

Frequent co-authors