Skip to content

Research

Two threads: a healthcare AI platform built on knowledge graphs and patient digital twins, and ADAPT-Q, a family of fine-tuning methods for specializing language models without making them forget.

Loom: healthcare AI platform

loom.clinic

The problem

Healthcare generates roughly 2.5 exabytes of data a day12, yet less than 3% of it is used for predictive analytics3, and the volume doubles about every 73 days4. The data is fragmented across incompatible systems5. Up to 85% of healthcare AI projects never reach production67, and bias remains a serious, under-addressed risk8. Above all, we rarely practice personalized medicine at scale. That would mean tailoring care to each patient's biology, history, and social context, informed by outcomes in cohorts of similar patients.

The approach

Drawing on my PhD in biochemistry and molecular biology, a fellowship in precision oncology, and many years building AI platforms in industry, I designed Loom around four ideas:

  1. Knowledge-graph patient intelligence. Model each patient within their relationships to conditions, treatments, providers, and similar patients. Graph neural networks over these structures have shown 15–30% accuracy gains over tabular approaches910, with interpretable paths for clinical reasoning.
  2. Privacy-preserving multi-site collaboration. Federated learning lets institutions train together without sharing patient data, which can match centralized performance111213. This targets the data silos that 73% of healthcare organizations cite as a major barrier14.
  3. Clinical workflow integration. Predictions come with explanations and uncertainty estimates. Explanations measurably raise clinician trust1516, and most decision-support tools see adoption below 15%1718. These clinical dashboards are called Workbenches.
  4. Research enablement. Biomarker discovery, disease-mechanism work, drug repurposing, and treatment-pathway optimization through patient digital twins192021.

Where it stands

  • Knowledge graph of 2.3M+ patients with 2.3B+ nodes and 4.8B+ relationships
  • Multimodal integration of 18+ biomedical data sources: 143M+ encounters, 135M+ clinical notes, 10.8M+ imaging studies, and 1.74B observations
  • 100K+ line Python codebase, built and operated end to end since 2006
  • AI infrastructure for graph neural networks, temporal modeling, and uncertainty quantification
  • Workbenches: 15+ clinical interfaces and 100+ service methods
  • End-to-end MLOps: model registry and observatory, evaluation pipelines, drift detection, and vector databases for RAG
  • Privacy-preserving federated learning across sites, using differential privacy and homomorphic encryption

Research write-ups

Applied research from the Loom program. Each card links to the full article on loom.clinic.

  • Meta-Objective Discovery

    AI agents that discover and evolve what to optimize, not just how. Validated at Mayo Clinic for GraphRAG optimization, where it surfaced new levers such as citation density and temporal recency.

  • Emergent Clinical Concepts

    Agent-driven discovery of new conceptual units in clinical text (implicit relationships, meaningful absences, temporal sequences), guided by clinical outcomes rather than semantic similarity.

  • Compositional Learning

    One model, many domains, without catastrophic forgetting: domain-specific adapters trained in orthogonal subspaces on a frozen base model, with router-based inference.

  • Production AI Systems at Scale

    A compositional architecture for continuous learning: real-time adaptation without forgetting, domain-specific compliance and governance, audit-ready multi-domain systems, and 10× lower infrastructure cost.

  • Agent Evaluation Reference Guide

    A structured reference for evaluating coding, conversational, research, RAG/GraphRAG, and computer-use agents: grader types, eval strategies, key benchmarks, and design principles.

Roadmap

Phase Focus
Advanced predictive intelligence Temporal GNNs for 1–12 month outcome prediction · digital-twin treatment simulation · federated pilot with health-system partners
Precision medicine at scale Wearable, environmental, and social-determinant data streams · graph-based biomarker discovery · automated trial design and patient matching

The long-term goal is the learning health system described by the National Academy of Medicine2223: one where every patient interaction improves care for the next patient.

ADAPT-Q

Specializing language models without making them forget.

ADAPT-Q is a family of fine-tuning methods aimed at the trade-off between domain specialization and general capability. The goal is to learn a demanding new domain as well as standard approaches do, while preserving substantially more of the model's general capability. We are evaluating it across several base models and four distinct professional domains, against established fine-tuning methods and forgetting-mitigation baselines.

Read more about ADAPT-Q

References

  1. IBM Institute for Business Value. "The healthcare data explosion: Improving analytics to drive innovation." IBM Corporation, 2020. ↩

  2. Stanford Medicine. "Health Trends Report: The Rise of the Data-Driven Physician." Stanford University School of Medicine, 2021. ↩

  3. Raghupathi, W., & Raghupathi, V. "Big data analytics in healthcare: promise and potential." Health Information Science and Systems, 2(1), 3, 2014. ↩

  4. Dinov, I.D. "Volume and Value of Big Biomedical Data." Journal of Medical Statistics and Informatics, 4(1), 2016. ↩

  5. Adler-Milstein, J., et al. "Electronic health record adoption in US hospitals: the emergence of a digital 'advanced use' divide." JAMIA, 24(6), 1142-1148, 2017. ↩

  6. Gartner. "Maverick Research: Healthcare's AI Challenge Is Not About Technology — It's About Trust." Gartner Research, 2022. ↩

  7. Sendak, M.P., et al. "A Path for Translation of Machine Learning Products into Healthcare Delivery." EMJ Innovations, 10(1), 19-00172, 2020. ↩

  8. Obermeyer, Z., et al. "Dissecting racial bias in an algorithm used to manage the health of populations." Science, 366(6464), 447-453, 2019. ↩

  9. Li, X., et al. "Graph neural network-based diagnosis prediction." Big Data, 8(5), 379-390, 2020. ↩

  10. Zhou, J., et al. "Graph neural networks: A review of methods and applications." AI Open, 1, 57-81, 2020. ↩

  11. McMahan, B., et al. "Communication-efficient learning of deep networks from decentralized data." AISTATS, 2017. ↩

  12. Rieke, N., et al. "The future of digital health with federated learning." NPJ Digital Medicine, 3(1), 119, 2020. ↩

  13. Sheller, M.J., et al. "Federated learning in medicine: facilitating multi-institutional collaborations without sharing patient data." Scientific Reports, 10(1), 12598, 2020. ↩

  14. Healthcare Information and Management Systems Society (HIMSS). "2022 Healthcare Data Analytics Survey." HIMSS Analytics, 2022. ↩

  15. Tonekaboni, S., et al. "What clinicians want: contextualizing explainable machine learning for clinical end use." Machine Learning for Healthcare Conference, 359-380, 2019. ↩

  16. Jacobs, M., et al. "How machine-learning recommendations influence clinician treatment selections: the example of antidepressant selection." Translational Psychiatry, 11(1), 108, 2021. ↩

  17. Sutton, R.T., et al. "An overview of clinical decision support systems: benefits, risks, and strategies for success." NPJ Digital Medicine, 3(1), 17, 2020. ↩

  18. Liberati, E.G., et al. "What hinders the uptake of computerized decision support systems in hospitals? A qualitative study and framework for implementation." Implementation Science, 12(1), 113, 2017. ↩

  19. Venkatesh, K.P., et al. "Digital Twins for Health: Opportunities, Challenges, and Practical Implications." Nature Medicine, 28(11), 2188-2190, 2022. ↩

  20. Laubenbacher, R., et al. "Using digital twins in viral infection." Science, 371(6534), 1105-1106, 2021. ↩

  21. Corral-Acero, J., et al. "The 'Digital Twin' to enable the vision of precision cardiology." European Heart Journal, 41(48), 4556-4564, 2020. ↩

  22. Institute of Medicine. "Best Care at Lower Cost: The Path to Continuously Learning Health Care in America." The National Academies Press, 2013. ↩

  23. Friedman, C.P., et al. "Toward a science of learning systems: a research agenda for the high-functioning Learning Health System." JAMIA, 22(1), 43-50, 2015. ↩