Machine learning for clinical decision support

Yanke Li

李衍锞

Postdoctoral Researcher in AI for Robust Clinical Prediction ModellingNuffield Department of Women's & Reproductive Health · Big Data Institute · University of Oxford

PhD, ETH Zürich · Spinal Cord Artificial Intelligence (SCAI) Lab & Sensory-Motor Systems Lab

I build the parts of an explainable clinical digital twin: models of individual patients that keep their state in a form a clinician can inspect, predict how it changes, and answer what-if questions about treatment. My work joins two causal traditions that are usually kept apart — structure learning with graphical models, and potential-outcomes reasoning — with deep temporal models that are interpretable by design rather than explained after the fact.

I completed my doctorate at ETH Zürich with Dr. Diego Paez-Granados, Prof. Robert Riener, and Prof. Julia Vogt, working closely with clinicians at the Swiss Paraplegic Centre. I am now in Prof. Christopher Yau's group at the University of Oxford, working on simulation-based pre-training for robust and trustworthy clinical prediction models.

Portrait of Yanke Li
Oxford, United Kingdom
Recent

News

  • Oct 2026

    I joined the University of Oxford as a Postdoctoral Researcher in AI for Robust Clinical Prediction Modelling, working with Prof. Christopher Yau in the Nuffield Department of Women's & Reproductive Health and the Big Data Institute.

  • Sep 2026

    I successfully defended my doctoral thesis, Towards Explainable Digital Twins for Clinical Decision Support across Spinal Cord Injury and Intensive Care, at ETH Zürich on 10 September 2026.

  • Jan 2026

    Two co-first-author papers accepted at ICLR 2026: GARLIC, on interpretable graph attention for intensive-care time series, and BioBO, on biology-informed Bayesian optimization for perturbation design.

  • Dec 2025

    KarmaTS was presented at ML4H 2025 in San Diego, supported by an ML4H Travel Award.

  • Oct 2025

    Oral presentation at the ISCoS Annual Meeting in Gothenburg on graphical modelling for pressure-injury risk prediction.

  • Sep 2025

    Completed a PhD research internship in Data Science & AI at Johnson & Johnson Innovative Medicine (June–September 2025), where BioBO was developed.

  • Apr 2025

    Our team placed first of 14 in the ASIA Data Science Challenge on predicting functional recovery after traumatic spinal cord injury.

What I work on

Research

A clinical record tells you what happened — never what would have happened. My research builds models that clinicians can question anyway: graphs that say which factors carry risk, temporal models whose explanation is the model itself, and simulation frameworks that manufacture the ground truth needed to check a causal claim.

  • RQ1represent · Part I
  • RQ2predict · Part II
  • RQ3act · Part III
  • RQ4personalise · ICare-DT
Injury Acute care Rehabilitation Home Follow-up day 0 days – weeks 3 – 6 months years decades PART I Snapshot one row per patient PART II Signal one stay, over time PART III Action effects & experiments structure shared zoom in: wound bed attention steering yardstick synthetic state · forecast · known truth candidate targets HAPI PI Fusion Microbiome GARLIC KarmaTS ICare-DT BioBO
  • a study, under the stage it addresses
  • how one study leads into the next
  • synthetic — no patient data
  • split colour — also answers RQ4, personalise
Graphicalmodels Causaldiscovery Foundationmodels Attention Causaleffects Datafusion causal graph + predictor fused graphs steer a sparse statistics + graph relational learning by executable causal process ICU FM + gated adaptation multimodal Bayesian optimisation one tabular cohort tabular FM · two cohorts 16S wound profiles graph attention · ICU synthetic ground truth + causal readout with a biology-informed prior HAPI PI Fusion Microbiome GARLIC KarmaTS ICare-DT BioBO
  • Part I · Snapshot
  • Part II · Signal
  • Part III · Action
  • a thread — the same method, study to study
Seven studies along the patient journey, each placed under the stage its question addresses (KarmaTS is synthetic, and BioBO's data are laboratory CRISPR screens); the arrows show how one study leads into the next. No study sits under follow-up yet, because no dataset covers those decades. The same seven studies, read by method: a dot marks the methods each study builds on or contributes to, and a thread joins the studies that share it. Two ideas run through all of them — explanation built into the structure, and causality.

Part I · Snapshot

Inspectable risk structure, one row per patient

HAPI method: mixed-type data from 250 patients are tested with PPCIT, a graph is learned with FCI and expert time tiers, its Markov blanket picks the features for an ensemble risk score, and the readable graph stays open to argument. Problem · method · results
Part I · SnapshotSci. Rep. 2024

HAPI

Risk flagged upstream — with a graph you can argue with

After spinal cord injury, a pressure injury becomes visible only once the damage is done, and today's risk forms give a score with no reason. Mixed-variable graphical modelling learns a causal graph of the risk factors from 250 patients, with clinicians' timeline orienting the arrows, and predicts from that graph. The readable graph costs nothing in accuracy; its edges are hypotheses to argue with, not facts.

  • AUROC 0.952 with the full Markov blanket (nested CV)
  • Admission-time features alone: balanced accuracy 0.925 (held-out split)
PI Fusion method: the HAPI and PPIS cohorts learn a shared graph that constrains each cohort-specific graph; re-imputation repeats, and the graph biases the attention of a frozen tabular foundation model. Problem · method · results
Part I · Snapshotnpj AI · submitted

PI Fusion

Two small cohorts — share structure, never parameters

Surgeons want to know what drives complications after a pressure injury is repaired, yet each cohort (250 acute-care patients; 266 surgical episodes from 174 patients) is too small to learn a reliable graph alone. Causal graph fusion shares structure, never parameters: a graph over the shared variables constrains each cohort's graph and steers a frozen tabular foundation model. A compact graph-selected feature set predicts complications about as well as all 75 features.

  • Age, albumin, haemoglobin → length of stay in both cohorts
  • Compact graph set ≈ all 75 features (held-out AUROC)
  • Graph F1 0.62 → 0.70 against synthetic known truth
Microbiome method: 55 wounds profiled by 16S rRNA sequencing, healed versus non-healing, read through three lenses: differential abundance, model attribution and co-abundance centrality. Problem · method · results
Part I · SnapshotSci. Rep. · under review

Wound microbiome

Same care, opposite outcomes — does the wound's community explain why?

Two pressure wounds can match in stage, size, site and care, yet months later one has improved and the other has stalled or worsened. Three analytically distinct methods, applied to 16S rRNA profiles of 55 wounds, test whether the bacterial community explains healing. All three agree: anaerobic genera are enriched in wounds that fail to improve, and a burden score over seven genera discriminates about as well as the full classifier, pending prospective validation.

  • Anaerobic burden ≈2× in poorly healing wounds, p = 0.002
  • Anaerobes are the network hubs (top hub degree 0.87)
  • 7-genus burden score ≈ gradient-boosted classifier
239 adults with SCI one camera body mesh scaled model side view virtual markers OpenSim fit drive hip knee ankle ground force r 0.68–0.90 r 0.85–0.87 vs motion capture vs force plates Markov blanket of functional independence
Related · RehabilitationarXiv 2026

Markerless gait

Few clinics have motion capture — gait biomechanics from a single video

Gait analysis guides rehabilitation after spinal cord injury, but it relies on motion capture and force plates that few clinics have. On the SCAI SCI Gait dataset (239 adults), a parametric body mesh fitted to a single side-view video drives a scaled OpenSim model. Markerless kinematics agreed with motion capture, predicted ground-reaction forces closely matched force plates, and Markov-blanket graph analysis found waveform components conditionally associated with functional independence.

  • Kinematics vs motion capture: r = 0.68–0.90
  • Ground-reaction forces vs force plates: r = 0.85–0.87
  • SCAI SCI Gait dataset: 239 adults with SCI

Part II · Signal

Forecasts that show their reasons, one stay over time

GARLIC method: gaps are treated as information, a time-lagged dependency graph shows which channel drives which, and cross-dimensional attention gives readable weights; the attention and edges are the explanation, with nothing added afterwards. Problem · method · results
Part II · SignalICLR 2026

GARLIC

ICU records are irregular — a forecast that shows its own reasons

Intensive-care records are irregular and up to 98 percent empty, and an explanation added afterwards need not be faithful. GARLIC treats gaps as information and learns a time-lagged dependency graph between signals; its attention weights and edges are themselves the explanation. It ranks first on three ICU benchmarks against 15 baselines, and its explanations pass a prespecified remove-and-retrain audit.

  • Best AUROC and AUPRC on 3 ICU benchmarks, 15 baselines
  • MIMIC-III AUROC 90.1 vs 89.2 for the runner-up
  • Remove-and-retrain audit: 8/8 checks vs 5/8 (RETAIN)
KarmaTS method: real series feed causal discovery and edge fitting, experts edit the lagged graph, and the executable process simulates synthetic series with a known graph, with feedback for expert refinement. Problem · method · results
Part II · SignalML4H 2025

KarmaTS

No ground truth for causal claims — so we built one

Causal claims on real data have no ground truth: known-truth benchmarks use random graphs, and real recordings hide the mechanism. KarmaTS makes a discrete-time causal process executable: starting from a real recording, an algorithm proposes the structure, an expert can edit it, and the result can be run forward, intervened on or shifted. Six discovery algorithms were graded against template graphs with a known answer; expert-elicited graphs and a user study are still to come.

  • 6 discovery algorithms graded on F1 and SID
  • Varies series length, lag, density and hidden variables
  • Its known-truth logic carries into ICare-DT's counterfactual benchmark

Part III · Action

Effects and experiments, graded against known truth

ICare-DT method: a transformer reads the patient history and outputs competing risks, trajectory forecasts, treatment effects and a personal safety check; the same history is queried under do(A=1) and do(A=0) and graded against a known effect on semi-synthetic eICU data; the model is fitted on eICU and tested for transfer on MIMIC. Problem · method · results
Part III · Actionnpj AI · submitted

ICare-DT

What would an action do? Effects graded, personal updates checked

Only one future is recorded, an average effect is not one patient's, and refitting on a few hours of one patient risks harmful drift. ICare-DT learns one representation of an ICU patient's history, with separately testable heads for risk, trajectory, treatment effect and gated personal updates. On semi-synthetic known truth it cut effect error over tenfold versus naive association; a personal update failed its prespecified check and was refused.

  • Treatment-effect error 0.008 vs 0.155 naive
  • Error 0.854 for RMSN, 0.929 for causal transformer
  • Gate vetoed update: 0 / 48 subgroup tests passed
BioBO method: a loop of biology prior, surrogate, acquisition and experiment, in which the surrogate fuses multimodal gene embeddings and a biology-informed acquisition prior decays as evidence accumulates. Problem · method · results
Part III · ActionICLR 2026

BioBO

Vast search space, tiny budget — prior biology must guide without deciding

A lab can afford only a few hundred single-gene experiments, under 5% of some 10,000 candidate genes, so prior biology must guide the search without deciding it. BioBO is a Bayesian optimisation loop whose surrogate fuses several data modalities and whose acquisition uses a biology-informed prior that fades as real data arrive. On GeneDisco CRISPR screens it finds the best genes sooner than standard Bayesian optimisation; a deliberately wrong prior falls back to standard performance.

  • 4–40% labelling effort saved by modality fusion
  • Biology prior improves the search in 23 of 24 settings
  • Wrong prior: 0.091 vs 0.093 for standard BO (IFN-γ)

Medicine will not be improved by models that take the decision away from the clinician. It will be improved by models a clinician can work with, and nobody can work with a model they cannot question.

— from my doctoral thesis, Towards Explainable Digital Twins for Clinical Decision Support across Spinal Cord Injury and Intensive Care (ETH Zürich, 2026)
Now at Oxford

Simulation-based pre-training for robust clinical prediction

I work in Prof. Christopher Yau's group, in the Nuffield Department of Women's & Reproductive Health and the Big Data Institute, on an MRC–NIHR funded project. Clinical prediction models can lose performance when the data change, and biased training data can make their predictions unfair. The project is developing a general framework that addresses both during training, by preparing a model for situations its data barely cover. It is designed to work across model types and clinical tasks, and has three parts:

  1. Simulated inputsExamples of challenging situations, such as rare patient profiles or data that have shifted.
  2. Desired behaviourHow the model should act on those inputs, such as staying stable under data drift or meeting a fairness constraint.
  3. Combined trainingThe model fits real outcomes as usual, with a penalty for departing from that behaviour.

The project will test this on three existing prediction problems:

  • Ovarian cancer prognosis: robustness
  • COVID-19 outcomes: data drift
  • Childhood mental health: fairness

I come to this from my PhD work on testing methods against known truth (KarmaTS) and guiding a search with prior knowledge (BioBO).

Papers

Publications

Published & accepted

Under review

  • ICare-DT: A Digital Twin for Continuous Forecasting of Intensive Care Patients npj Artificial Intelligence Y. Li, Y. Qi, Y. Kim, D. Paez-Granados
  • Causal Graph Fusion Across Incomplete Cohorts and Causal Steering of Tabular Foundation Models npj Artificial Intelligence Y. Li, D. Al Jorf, A. Ighissou, A. Scheel-Sailer, R. Riener, D. Paez-Granados
  • Anaerobic Microbial Communities Influence Wound Recovery Outcome in Pressure Injuries: A Machine Learning Approach Scientific Reports Y. Li, J. Stoyanov, R. Riener, R. Wettstein, S. Capossela, D. Paez-Granados†, A. Bertolo†
  • Patient Digital Twins for Caregiving Robots Book chapter D. Paez-Granados, Y. Li, M. Ejtehadi, B. Fuchs, O. Gnarra (Springer Moonshot Consortium)

Preprints

  • Monocular Markerless Biomechanics for Clinically Interpretable Gait Assessment in Spinal Cord Injury arXiv 2026 S. Natraj, M. Ruepp, Y. Li, R. Riener, I. Eriks-Hoogland, D. Paez-Granados · arXiv:2610.05552

In preparation

  • Collaborative Causal Discovery with Multimodal Data in Spinal Cord Injury Y. Li*, S. Zhong*, H. Li, R. Riener, D. Paez-Granados
  • GRITS: Graphical Imputation for Multivariate Time Series S. Joray, Y. Li, D. Paez-Granados

* equal contribution · † shared senior authorship

Presentations

Talks & posters

  • Apr 2026 ICLR 2026 — GARLIC and BioBO (posters) Rio de Janeiro, Brazil
  • Dec 2025 Machine Learning for Health (ML4H) — KarmaTS (poster) San Diego, United States
  • Oct 2025 ISCoS Annual Meeting — Mixed-variable graphical modelling towards risk prediction of hospital-acquired pressure injury (oral) Gothenburg, Sweden
  • May 2024 AI for Good Global Summit — A digital twin framework for chronic diseases (booth) Geneva, Switzerland
  • Apr 2023 European Causal Inference Meeting — Continual causal discovery and inference for SCI comorbidities (poster) Oslo, Norway
Academic service

Teaching, mentoring & review

Teaching

Fundamentals of Artificial Intelligence Centre for Doctoral Training (FoAI CDT), University of Oxford
Teaching Assistant — 2026–present

Artificial Intelligence for Healthcare and Rehabilitation, ETH Zürich
Teaching Assistant — Autumn 2024 & Autumn 2025

Peer review

Conferences: NeurIPS, ICLR, ICML, AISTATS, ML4H

Journals: Journal of NeuroEngineering and Rehabilitation

Student supervision

  • 2026D. Al Jorf — Master's lab practice, ETH Zürich
  • 2025R. Wang — Master's thesis, University of Zurich
  • 2025Y. Qi — Master's thesis, ETH Zürich
  • 2025S. Joray — Master's internship, ETH Zürich
  • 2024A. Ighissou — Master's thesis, University of Milano-Bicocca
  • 2024J. Koch — Semester project, ETH Zürich
  • 2023H. Li — Master's internship, TU Munich
  • 2023B. Wu — Master's thesis, TU Munich
Recognition

Awards & scholarships

  • Nov 2025

    Machine Learning for Health (ML4H) Travel Award

  • Apr 2025

    First place (of 14 teams), ASIA Data Science Challenge — functional recovery prediction after traumatic spinal cord injury

  • Feb 2018

    Outstanding Student Award, Faculty of Applied Science & Textiles, PolyU — top graduate of the faculty; BSc GPA 4.00/4.00

  • 2014–18

    Dean's Honours List (top 5%), Faculty of Applied Science & Textiles, PolyU · Ho & Ho Foundation Entry Scholarship for outstanding students

  • May 2017

    Bauhinia Cup Outstanding Entrepreneurs Association Scholarship

  • Oct 2016

    Commercial Radio 50th Anniversary Scholarship

  • Jun 2015

    Reaching Out Award, Hong Kong Government Scholarship