Preventative Health Bench01
The endpoint belongs in the result name
Published evidence
The study changes endpoints and predictor sets explicitly. [2]
Our interpretation
Our inference: an algorithm ranking without its endpoint and available features is underspecified.
Independent benchmark analysis / 2026 accepted journal manuscript
UK Biobank survival-model benchmark by Oexner et al.
Ranking future risk is only one part of a prevention decision.
This published benchmarking study compares survival models across incident cardiovascular disease, breast cancer and Alzheimer’s disease, using several feature regimes. It supplies a concrete prevention-related evaluation rather than a general screening checklist. Our analysis asks how the endpoint, available information and validation design shape a model comparison. The reported outcome is risk discrimination: which people are ranked ahead of others. It does not by itself establish calibrated absolute probabilities or the benefit of an intervention. We retain the study’s name, conditions and uncertainty rather than branding its data as a new benchmark of our own.
01 / What is being tested?
The task, before the score.
- input
- Baseline participant predictors under a specified feature regime
- output
- A survival-risk score for a defined incident-disease endpoint
- unit
- One participant with event/censoring time
- setting
- Nested five-fold cross-validation; no single fixed public test set
Data origin. Longitudinal UK Biobank participant information and linked health outcomes; access-controlled research data. [1][2][3]
- Disease endpoints
- 3
Cardiovascular disease, breast cancer and Alzheimer’s disease.
§2.2 [2] - Feature regimes
- 5
Age/sex; clinical risk; PRS/metabolomics; medical history; complete.
§2.3 [2] - Primary model families
- 8
Cox, Lasso, Ridge, Elastic Net, random forest, LightGBM, XGBoost and deep learning.
Model implementations [2] - Validation
- Nested 5-fold CV
Inner tuning is separate from outer evaluation; approximate cohort counts vary after endpoint exclusions.
§2.7 [2] - Analysis population
- Approximately 240k
Selected for <20% missingness across predictor matrices, largely requiring metabolomics coverage; before endpoint-specific baseline-disease exclusions.
§§2.4–2.5 [2]
- 01
Define incident outcomes
Exclude baseline occurrence of the corresponding endpoint.
[2] - 02
Specify available predictors
Fix one of the five feature regimes before comparison.
[2] - 03
Tune inside training folds
Use nested cross-validation and fitted preprocessing.
[2] - 04
Evaluate held-out ordering
Report fold-level discrimination and uncertainty for that endpoint.
[3][2]
02 / Measurement
Harrell’s concordance index
Higher is better
Measures ordering of risk against observable event order among comparable pairs; the paper also reports Uno’s C.
Scoring definition
C = concordant comparable pairs / comparable pairs, with the implementation’s tie handling
Compare identical endpoint, predictor regime and validation folds. C is not percent of patients correctly diagnosed. [2]
03 / Measured evidence
Results, with their conditions attached.
Paper-reported results / selected rows
Cardiovascular disease with clinical-risk features
2026 accepted manuscript and Supplementary Table 3; nested CV test performance for the Clinical Risk regime.
Harrell’s C · C-index
00.51
ReportedMinimally penalised Cox-PH95% CI 0.717–0.721.
0.719Deep learning95% CI 0.719–0.722.
0.721XGBoost95% CI 0.560–0.569.
0.564Study-specific implementations and tuning grids. These are discrimination values, not clinical accuracy percentages or universal algorithm rankings.
Source: Cardiovascular Disease worksheet; Clinical Risk rows [3][2]
04 / Our original analysis
What follows from the design?
02
Discrimination leaves an unanswered prevention question
Published evidence
The study focuses on discrimination rather than calibration. [2]
Our interpretation
Our inference: an ordering result cannot alone support an absolute-risk treatment threshold.
03
Validation units are participants, not pair counts
Published evidence
The C-index compares observable pairs constructed from participant outcomes. [2]
Our interpretation
Our inference: the number of pairs should not be used as though it were the number of independent patients when discussing uncertainty.
04
A small score gap is context dependent
Published evidence
The selected clinical CVD estimates differ by 0.002 between deep learning and Cox-PH. [3]
Our interpretation
Our arithmetic: 0.721 minus 0.719 equals 0.002; that difference is not a measured reduction in disease incidence.
05 / Scope of the evidence
Where this benchmark stops.
Selected source cohort and assay coverage
UK Biobank recruitment is selective. The analysis further filters for predictor availability and missingness, largely removing participants without metabolomics. Transport to another population needs separate evidence. [2]
No intervention endpoint
The comparison predicts disease events; it does not randomize or test a prevention policy. [2]
Controlled data access
The paper is open, but participant records require UK Biobank researcher approval. [2]
Accepted manuscript version
The journal labels this an article in press; the cited supplement is the 2026 version, not the 2025 preprint. [1]
- Availability
- Paper and aggregate supplements public; participant data restricted
- License
- Article CC BY 4.0; participant data under UK Biobank access terms
- Conditions
- Approved-researcher access is required for participant data. Article licensing does not grant dataset access.
Evidence trail
Read the originals.
- Comprehensive benchmarking of machine learning methods for risk prediction modelling from large-scale survival data: a UK Biobank study ↗
Oexner et al. / King’s College London and UCL. Accepted peer-reviewed article; results differ in some uncertainty estimates from the 2025 preprint.
- UK Biobank survival benchmark: accepted manuscript ↗
Oexner et al. / Journal of Big Data. Complete methods, endpoints, nested cross-validation and limitations.
- UK Biobank survival benchmark: Supplementary Table 3 ↗
Oexner et al. / Journal of Big Data. Published workbook containing condition-specific model results and confidence intervals.
Benchmark authors retain authorship of their work. This publication provides independent analysis; published rows are not Arcophos evaluation runs. Editorial method.