Jiayuan Xu Medical ML researcher

PhD researcher · University of Manchester

Machine learning for earlier, more accessible neuro­degenerative disease detection.

I build rigorous and clinically interpretable models from plasma biomarkers, genetic information and real-world clinical data—then test whether they generalise beyond the cohort they were trained on.

PhD completion expected Q4 2026 · Open to research and applied ML roles
Explore my research Read web CV

Research position

A high AUC is not enough.

Medical AI has to survive missing data, cohort shift, label design and clinical scrutiny. My work is centred on the evidence behind the score: leakage-free evaluation, independent external testing, honest limitations and explanations that domain collaborators can interrogate.

External validation Leakage-safe pipelines Methodological honesty Clinical interpretability

Selected research

Three questions, one translational direction.

From identifying brain amyloid without a PET scan, to forecasting disease progression, to checking whether diagnostic labels give a model an unfair shortcut.

01 · Amyloid pathology

Can a blood test triage who may need amyloid PET?

Built and externally evaluated an end-to-end pipeline using plasma biomarkers, APOE genotype and clinical information. Four algorithms were compared, with cross-cohort feature matching and SHAP-based interpretation.

Cohorts
ADNI + independent CNTN
Models
DT · RF · SVM · MLP
Contribution
Feature selection + zero-shot external transfer
Read the paper Inspect the code
ROC–AUC Amyloid-β PET positivity
Internal ADNIRF · n=340
0.95
External CNTNMLP · n=127 · no retraining
0.90
Reduced model5 features · MLP
0.87

The independent-cohort result is the key evidence: performance transferred without retraining.

Xu et al. · Frontiers in Aging Neuroscience · 2025

02 · Cognitive staging

What happens when the label already contains the answer?

Quantified label circularity in three-class CN/MCI/AD staging. Clinical scales used to define ADNI diagnoses produced near-ceiling scores; after removing that shortcut, the blood-based estimate was lower—but more credible.

Design
4 classifiers × 3 feature sets
Evaluation
Repeated nested CV · 15 outer folds
Transfer
Reduced-panel zero-shot evaluation in CNTN
Read the paper Reproduce the analysis
AUC–OVR Honest staging benchmarks
Clinical-onlycircularity-inflated upper bound
0.954
Full fusiongain vs clinical: p=0.33
0.956
Non-circular internalplasma + demographic + genetic
0.746
External CNTNreduced feature panel · n=130
0.702

The contribution is not a bigger number. It is a realistic estimate once definitional leakage is made explicit.

Xu & Costen · Diagnostics · 2026

03 · Disease progression

Can non-invasive biomarkers forecast MCI-to-AD conversion?

A voting ensemble combined logistic regression, random forest and LightGBM for two- and three-year progression prediction in 767 ADNI participants. Every imputation, scaling and resampling step was fitted inside the cross-validation fold.

ADI 2026 oral presentation
  1. Outer5-foldheld-out evaluation
  2. Inner3-foldhyperparameter search
  3. In-foldKNN · scale · SMOTEzero preprocessing leakage
3-year full-model AUC 0.888 Base + plasma: 0.881 · Base + CSF: 0.875

Earlier work · digital health

Two ways to observe movement—wearable and contactless.

A

Wearable sensing · 2023

Wearable pressure-map motion classification

Built an end-to-end system spanning ReTiSense’s Stridalyzer PRISM pressure-sensing insoles, Android/web data acquisition and a signal-to-image ML pipeline. Benchmarked four architectures on 12,000 pressure maps across six controlled motion classes, working directly with ReTiSense’s engineering team; this was an engineering benchmark rather than patient-level diagnostic validation.

  • Wearable sensors
  • Computer vision
  • LeNet–SVM
  • MATLAB
B

Edge sensing · 2022

Privacy-oriented thermal fall-detection prototype

Designed an offline, explainable thermal-imaging pipeline using non-linear Difference-of-Gaussians processing and a lightweight temporal-memory algorithm. The prototype avoids identifiable RGB imagery and cloud transfer.

  • Thermal imaging
  • Edge AI
  • Classical CV
  • Privacy by design

Selected talks & posters

Research presented—and debated—in the room.

Oral, panel and poster presentations on Alzheimer’s disease progression, plasma biomarkers and clinically credible machine learning.

2026 · Lyon, France · Oral presentation & panel

The Biomarker Revolution: Research and Practice

Selected as the session’s sole PhD-stage speaker, I presented a leakage-safe ensemble machine-learning framework for two- and three-year MCI-to-Alzheimer’s disease conversion prediction. I then joined the four-speaker panel as its sole AI/ML speaker, discussing how validated plasma biomarkers can support clinically relevant prediction tools.

  • AbstractP253
  • Session roleSole PhD-stage speaker
  • Panel roleSole AI/ML speaker
View the progression-prediction study
Jiayuan Xu presenting research on MCI-to-Alzheimer’s disease conversion at ADI 2026 in Lyon
Oral presentationPresenting ensemble machine learning for MCI-to-AD conversion prediction.
Jiayuan Xu speaking into a microphone during the ADI 2026 biomarker panel discussion
AI/ML perspectiveDiscussing how machine learning can bridge validated plasma biomarkers and clinical prediction.
Four speakers and the session chair in the ADI 2026 panel, with Professor Charlotte Teunissen at far left and Jiayuan Xu at far right
Panel discussionProf Charlotte Teunissen at far left; Jiayuan Xu at far right.
From left, Jiayuan Xu, Professor Linus Jönsson, Professor Séverine Sabia, Professor Charlotte Teunissen and session chair Haythum Tayeb at ADI 2026
Speakers & chairLeft to right: Jiayuan Xu, Prof Linus Jönsson, Prof Séverine Sabia, Prof Charlotte Teunissen and chair Haythum Tayeb.

Selected poster presentations

From diagnostic shortcuts to disease progression.

Preview of the AAIC 2026 poster on blood-based Alzheimer’s disease staging and label circularity
Jiayuan Xu beside his blood-based Alzheimer’s disease staging poster at AAIC 2026 in London
Poster presentation · London

AAIC 2026 · London · Poster

Blood-Based Staging of Alzheimer’s Disease—and the Label Circularity That Inflates It

Quantified how diagnostic-label circularity can inflate apparent model performance, then evaluated a non-circular plasma, demographic and genetic feature panel internally in ADNI and externally in CNTN.

Poster PDF ↗ Paper ↗ Code ↗
Preview of the AD/PD 2026 poster on six-year MCI-to-Alzheimer’s disease conversion prediction
Jiayuan Xu beside his MCI-to-Alzheimer’s disease conversion poster at AD/PD 2026 in Copenhagen
Poster presentation · Copenhagen

AD/PD 2026 · Copenhagen · Poster

An Ensemble Machine Learning Model for the Prediction of MCI-to-AD Conversion Within Six Years

An earlier multimodal ensemble-learning study of six-year MCI-to-AD conversion, preceding the stricter nested-cross-validation framework presented at ADI 2026.

Poster PDF ↗ Later ADI study ↑

Conference archive

  1. 2026 · Lyon
    Predicting MCI-to-AD Conversion Using Ensemble Learning and Multi-Modal BiomarkersOral presentation & panel · Alzheimer’s Disease International (ADI)
  2. 2026 · Copenhagen
    An Ensemble Machine Learning Model for the Prediction of MCI-to-AD Conversion Within Six YearsPoster · AD/PD
  3. 2026 · London
    Blood-Based Staging of Alzheimer’s Disease—and the Label Circularity That Inflates ItPoster · Alzheimer’s Association International Conference (AAIC) · Related paper ↗
  4. 2025 · Vienna
    Estimation of Amyloid-β PET Status Using Plasma Biomarkers and Clinical InformationPoster · AD/PD · Related paper ↗

Selected publications

Peer-reviewed work, with the code beside it.

Two first-author open-access journal articles, with public analysis code and links to their related conference outputs.

2026
Journal article · First author

Machine Learning-Based Multiclass Classification of Cognitive Stages Using Plasma Biomarkers, Clinical Assessments, and Genetic Features

Jiayuan Xu & Fumie Costen · Diagnostics 16(12), 1755

DOI ↗Code ↗AAIC poster ↑
2025
Journal article · First author

Accurate and Robust Prediction of Amyloid-β Brain Deposition from Plasma Biomarkers and Clinical Information Using Machine Learning

Jiayuan Xu, Andrew J. Doig, Sofia Michopoulou, Petroula Proitsi & Fumie Costen · Frontiers in Aging Neuroscience 17, 1559459

DOI ↗Code ↗AD/PD 2025 ↑

Experience

Research shaped by engineering, translation and teaching.

I began in automation and embedded systems, moved through robotics and wearable sensing, and now work at the intersection of biomarkers, clinical cohorts and machine learning.

2023—present

University of Manchester

Doctoral Researcher · Electrical & Electronic Engineering

Machine learning for Alzheimer’s prognosis, amyloid pathology and blood-based cognitive staging. PhD completion expected Q4 2026.

2023

De Montfort University · Industry project with ReTiSense

Graduate Researcher · Wearable Digital Health

End-to-end pressure-map motion-analysis system using Stridalyzer PRISM sensor insoles, from mobile data acquisition to ML evaluation.

2022

De Montfort University

Research Intern · Edge AI & Computer Vision

Privacy-oriented, contactless fall-detection prototype using a single thermal camera.

2024—2026

University of Manchester

Teaching Assistant · Digital Medical Image Processing

Supported the module and mentored more than ten MSc research projects spanning medical imaging and healthcare AI.

Education

  • PhD, Electrical & Electronic EngineeringUniversity of Manchester · expected Q4 2026
  • MSc, Intelligent Systems & RoboticsDe Montfort University · Distinction, 81/100 · 2023
  • Exchange StudentPurdue University Northwest · GPA 4.0/4.0 · 2020–2021
  • BEng, AutomationNanjing University of Science and Technology Zijin College · 2020

Recognition & service

  • Best Postgraduate Student PrizeDe Montfort University · 2023
  • Dean’s List & Semester HonorsPurdue University Northwest · 2021
  • Jiangsu Provincial Second PrizeBlue Bridge Cup C/C++ Programming · 2019
  • Ad hoc reviewerIET Image Processing · IEEE Transactions on Radiation and Plasma Medical Sciences

Capabilities

Methods before buzzwords.

Clinical ML methodology

Nested and repeated cross-validation · external cohort validation · ensemble learning · missing-data pipelines · class imbalance · feature matching · SHAP · ablation and sensitivity analysis

Programming & tooling

Python · PyTorch · scikit-learn · XGBoost · LightGBM · pandas · NumPy · SciPy · Matplotlib · Git · Linux · Docker · LaTeX · MATLAB

Biomedical data & domain

Plasma biomarkers · APOE genotype · ADNI and CNTN cohorts · Alzheimer’s progression · ATN framework · wearable sensors · pressure maps · thermal imaging · digital health

Contact

Let’s make clinical machine learning more credible—and more useful.

I am interested in research scientist, applied scientist and clinical data science roles spanning computational biomarkers, neurodegeneration, patient stratification and digital measures.

jiayuan.xu@manchester.ac.uk