ProjectQuintessence

Research

Clinical AI research from data construction to model training, evaluation, and deployment.

Project Quintessence builds clinical AI systems across language, agents, prediction, imaging, and physiologic modelling. The work spans dataset construction, supervised fine-tuning, reinforcement learning, encoder models, multimodal systems, forecasting, and clinically grounded evaluation.

01 / Clinical Foundation Models & Language Systems

Language Models & Clinical NLP

Clinical language models and NLP systems from data construction to expert-led validation.

Signal, methods & studies

Signal

Notes + guidelines + tool use

Validation principle

We evaluate models by clinical correctness, evidence traceability, calibration, robustness, and expert adjudication. Fluency is not treated as evidence of competence.

Methods

  • data construction
  • SFT
  • RL / preference optimization
  • encoder models
  • retrieval grounding
  • expert evaluation

Outputs

  • post-trained LLMs
  • BERT-based clinical classifiers
  • SFT and preference datasets
  • expert-validated benchmarks
  • clinical reasoning evaluations
  • synthetic data generation

Projects (1)

NephroGPT: A Tool-Augmented Large Language Model for Nephrology Training and Continuing Education

Validation planning

A tool-augmented language-model approach to specialist education and evidence-grounded clinical reasoning.

02 / CLINICAL AGENTS WITH AUDITABLE TOOL USE

Agentic AI Systems

Specialist agent workflows with auditable tool use and evidence-grounded decision traces.

Signal, methods & studies

Signal

Evidence + context + tools

Validation principle

Agentic performance is accepted only when plans, tool calls, evidence inputs, intermediate states, and failure boundaries are inspectable by clinical reviewers.

Methods

  • agent orchestration
  • tool-use policy learning
  • workflow state modelling
  • retrieval / EHR grounding
  • multimodal context fusion
  • agent trajectory evaluation

Outputs

  • specialist agent workflows
  • auditable tool-execution traces
  • workflow simulation benchmarks
  • clinical task routers
  • safety and failure analyses

Projects (1)

An Evidence-Grounded Multimodal Agentic AI Co-Pilot for Cardiologists

Ongoing

An agentic system that integrates multimodal context and external tools to support cardiology decision workflows.

03 / TEMPORAL PREDICTION AND POPULATION DYNAMICS

Predictive Modelling & Forecasting

Risk estimation, temporal modelling, intervention analysis, and reproducible outcome validation.

Signal, methods & studies

Signal

Surveillance + interventions + time

Validation principle

Prediction and forecasting models are evaluated by temporal holdout performance, calibration, drift sensitivity, intervention timing, and reproducibility across cohorts—not one-off fit metrics.

Methods

  • risk prediction modelling
  • time-series forecasting
  • longitudinal representation learning
  • survival / event modelling
  • causal intervention analysis
  • calibration and drift monitoring

Outputs

  • validated risk models
  • forecasting pipelines
  • intervention impact estimates
  • temporal surveillance analyses
  • reproducible prediction benchmarks

Projects (1)

The Impact of COVID-19 Non-Pharmaceutical Interventions on Notifiable Infectious Diseases in Poland

Completed

A completed study on how pandemic-era interventions affected notifiable infectious disease trends in Poland from 2014 to 2022.

04 / MULTIMODAL CLINICAL VISION

Computer Vision & Medical Imaging

Medical imaging models spanning chest radiography, report-derived supervision, and multimodal post-training.

Signal, methods & studies

Signal

CXR + reports + biomarkers

Validation principle

Medical vision models are evaluated against image findings, report agreement, calibration, subgroup robustness, and downstream clinical meaning; benchmark scores are not treated as sufficient validation.

Methods

  • image-text alignment
  • report-derived supervision
  • multimodal post-training
  • foundation CXR modelling
  • biomarker validation
  • radiology benchmark curation

Outputs

  • CXR foundation model variants
  • automated report labelers
  • image-report datasets
  • imaging-derived biomarkers
  • clinical vision benchmarks

Projects (3)

Benchmarking Large Language Models for Automated CheXpert-Style Labeling of Chest X-Ray Reports

Completed

ChexQwen: A Post-Trained Multimodal LLM as a Foundation CXR Model

Completed

Chest X-Ray-Predicted Age as a Biomarker for Mortality

Completed

05 / LATENT PHYSIOLOGIC DYNAMICS

Physiologic World Models

Learned models of patient-state dynamics that forecast trajectories and simulate counterfactual futures.

Signal, methods & studies

Signal

Vitals + labs + synthetic signals

Validation principle

A physiologic world model is valid only if its learned state dynamics preserve temporal coherence, cross-modal consistency, intervention sensitivity, and calibrated uncertainty against held-out clinical trajectories.

Methods

  • latent state-space modelling
  • temporal representation learning
  • cross-modal generation
  • counterfactual simulation
  • uncertainty quantification
  • trajectory validation

Outputs

  • patient-state representations
  • physiologic trajectory simulators
  • ECG-to-echo generation systems
  • multimodal surrogate models
  • uncertainty-calibrated synthetic cohorts

Projects (1)

EchoWorld: Generative Echocardiography from ECG and Clinical Context

Ongoing

A physiologic-to-imaging generative study for 9-view echocardiography synthesis with uncertainty-aware outputs.