Bhaskarjyoti Das

Research profile

Research summary

Bhaskarjyoti Das is a researcher at PES University whose cataloged work spans two distinct lines: clinical machine learning for anaesthesia informatics and machine learning methodology for causal discovery. Both entries in Evolve's catalog are 2026 arXiv preprints with Das in the last author position, indicating a senior or supervisory role on small teams rather than a corresponding-author record. The clinical line contributes pharmacokinetic state space models for unbiased prediction of haemodynamic collapse, combining drug effect-site concentrations with a Mamba-style architecture and lead-gap filtering to report AUROC 0.736 for intraoperative hypotension prediction on the VitalDB dataset; Evolve's verdict credits solid ablations while cautioning that mechanism claims overreach the evidence. The methods line, 'Guide, Not Bind', shows through suppression and an identifiability tie why defeasible priors fail in augmented Lagrangian causal discovery — a synthetic-only study praised for honest limits though modest in practical scope. The field distribution across cataloged papers splits evenly (0.5/0.5) between the two areas, which yields a focus score of 0, i.e. low topical concentration. Reproducibility averages 78 across 2 scored papers (high readiness for the causal-discovery work, medium for the clinical model). Collaboration scores 39, with 6 distinct co-authors including Rithin Nagaraj, Sudiksha Chindula and Sairam Sundararaman. Output, impact and activity axes are not yet scored due to insufficient bibliometric data.

Research keywords

Highlights

• Pharmacokinetic state-space models for intraoperative hypotension prediction

  As last author, Das's team couples drug effect-site concentrations with a Mamba-style state space model and lead-gap filtering, reporting AUROC 0.736 for intraoperative hypotension prediction on VitalDB. Evolve credits the solid ablations but flags the mechanism claims as overreaching the evidence; reproducibility readiness is rated medium.

• Where defeasible priors fail in augmented Lagrangian causal discovery

  In a synthetic-only study, the paper's own defeasible augmented-Lagrangian rule is shown to fail through prior suppression and an identifiability tie, clarifying when forbidden-edge constraints can genuinely guide differentiable DAG learning. Evolve notes honest limits and modest practical scope; reproducibility readiness is high.

• A deliberately two-sided 2026 portfolio

  Both 2026 preprints place Das as last author yet target different communities: anaesthesia informatics and causal discovery methodology split the field distribution evenly at 0.5 each, with reproducibility readiness rated medium for the clinical model and high for the methods paper.

Timeline

1. 2026Publishes (as last author) 'Pharmacokinetic State Space Models for Unbiased Prediction of Haemodynamic Collapse' on arXiv, reporting AUROC 0.736 for intraoperative hypotension prediction on VitalDB.
2. 2026Publishes 'Guide, Not Bind: Why Defeasible Priors Fail in Augmented Lagrangian Causal Discovery' on arXiv, demonstrating failure via prior suppression and an identifiability tie.
3. 2026The two preprints land in the same year with an even 0.5/0.5 split between anaesthesia informatics and causal discovery, establishing the current dual-line profile.

Papers in the Evolve catalog

• 2026-09-21

  Pharmacokinetic State Space Models for Unbiased Prediction of Haemodynamic Collapse

• 2026-09-03

  Guide, Not Bind: Why Defeasible Priors Fail in Augmented Lagrangian Causal Discovery

Frequent co-authors