Can a biopsychosocial risk score stratify disease vulnerability in healthy people?

A score for cumulative burden, not one disease

Most risk tools answer a narrow question: will this person develop a particular condition? This paper asks a broader one. It defines disease vulnerability as the tendency to accumulate future adverse health outcomes, and it builds a single number, the Risk Score for Disease Vulnerability (RS4DV), to rank ostensibly healthy people on that tendency.

The target the model learns is cumulative disease burden: a simple count of diagnosed conditions from a predefined list, drawn from ICD-10 codes and recorded before the baseline assessment. That list mixes physical conditions such as hypertension, diabetes, asthma and arthritis with mental disorders such as depression, anxiety disorders and schizophrenia. The authors state they did not weight conditions by clinical severity; the count is meant as a general index of health status, not a comorbidity measure.

This choice shapes what the score can and cannot mean. A high RS4DV indicates a person whose measured profile resembles those who later accumulated many diagnoses. It does not name which disease will come next, and the paper explicitly notes that RS4DV does not precisely predict the exact number of future diseases for an individual.

Eighty-five measures, one gradient-boosted model

The inputs are 85 variables collected at the baseline touchscreen assessment, grouped into three domains. Mental health covers neuroticism items, recent trauma events, and emotional factors including consultations for nerves, anxiety, tension or depression. Physical health covers activity measured by MET score, sleep questions, smoking and alcohol use, and anthropometric measures such as BMI and blood pressure. Sociodemographic factors cover education, income, employment, household composition and demographics.

Participants with more than 20% missing values across these features were excluded, leaving 482,057 of 502,180. Remaining missing values were filled by Bayesian ridge regression imputation, performed separately within the discovery and test cohorts to avoid information leakage. The discovery cohort of 391,193 was split 7:3 into training and validation sets, and a held-out test cohort of 90,864 was reserved for evaluation.

Five algorithms were compared on training and validation data using R², mean squared error and mean absolute error, and Light Gradient Boosting Machine was selected to build the final model. Hyperparameters were tuned by grid search with five-fold cross-validation, and validation performance was used for early stopping to limit overfitting. The held-out test cohort played no role in algorithm selection.

Risk separation held over a median 14.7 years

The prospective test that matters most used held-out participants with no baseline diagnosis: 35,193 in the test set, alongside 46,550 in the internal validation set. Over a median follow-up of 14.7 years, baseline RS4DV tracked later disease accumulation, with correlations of ρ = 0.30 in the test subset and ρ = 0.32 in the validation subset.

The gap between risk groups was large. By year 15 in the test set, cumulative incidence of disease progression was 62.9% in the top RS4DV decile versus 21.5% in the bottom decile; cumulative all-cause mortality was 11.2% versus 1.0%. In the validation set the corresponding figures were 64.5% versus 18.4% for progression and 13.3% versus 1.6% for mortality.

In univariable Cox models, each one-unit increase in RS4DV was associated with disease progression (HR = 2.60, 95% CI 2.51–2.70 in the test set) and all-cause mortality (HR = 4.03, 95% CI 3.68–4.41). These are associations from an observational cohort, not evidence that changing the score would change outcomes. Disease-specific analyses found higher incidence in the top decile for 25 diseases, with cardiometabolic and mental health disorders predominating among high-risk individuals.

Modest gains beyond frailty and conventional factors

A score that merely re-expresses age or existing frailty would add little. The paper benchmarks RS4DV against two established measures. As a standalone predictor in the held-out test set, it reached a C-index of 0.620 for disease progression versus 0.576 for the frailty index, and 0.703 versus 0.598 for mortality. Against the proteomic biological age gap, it reached 0.627 versus 0.507 for progression and 0.685 versus 0.578 for mortality.

The incremental gains were smaller. Adding RS4DV to conventional covariates raised the test C-index from 0.611 to 0.626 for progression and from 0.764 to 0.771 for mortality. Adding it to models already containing conventional covariates and frailty raised the C-indices from 0.627 to 0.632 and from 0.771 to 0.774. In the proteomic subset, adding it on top of covariates, frailty and the proteomic age gap moved the C-index from 0.641 to 0.645 for progression and from 0.765 to 0.767 for mortality.

So the honest reading is a score with moderate discrimination that carries some information not captured by deficit accumulation or proteomic ageing, but whose marginal contribution on top of a full model is small. Calibration was close to observed risks at 10 and 15 years, and decision-curve analysis indicated only modest additional net benefit over conventional covariates alone.

Genomic, proteomic and imaging signals converge

To ask whether the score reflects biology rather than questionnaire artefact, the authors examined three molecular and imaging layers. A genome-wide association study among 127,898 disease-free individuals identified 10 genome-wide significant loci mapping to 26 genes, enriched in pathways for MAPK cascade regulation, DNA repair and intracellular protein transport. Genomic inflation was modest (λ = 1.25), and LD score regression put the intercept at 1.03, which the authors read as consistent with polygenicity rather than population stratification.

A polygenic risk score derived from that GWAS correlated only weakly with RS4DV at baseline (ρ = 0.06) but showed a graded pattern across disease-progression groups over follow-up. Proteomic and neuroimaging analyses likewise identified correlates, and mediation models linked polygenic risk, protein groups and brain structure to RS4DV.

These analyses support biological plausibility across organisational levels, which is the authors' interpretation. They do not establish that the psychosocial or lifestyle inputs cause later disease, and the mediation results come from observational modelling rather than intervention.

Six routine items retained much of the ranking

For practical use, the authors reduced the model to six routinely accessible clinical and self-reported measures spanning physical, psychological and general health domains. The full model had highlighted mental-health consultation, age, anthropometric measures including BMI and grip strength, systolic blood pressure, and tiredness or lethargy as important predictors.

The simplified model showed moderate predictive performance and high concordance with the full model. Its risk separation persisted: by year 15 in the test subset, disease progression was 60.3% in the top decile versus 22.3% in the bottom decile, and mortality was 9.0% versus 1.0%. For disease progression, the six-item score separated risk better than any of its six individual components; for mortality it was comparable to age and better than the other five.

Transferability was tested in the 1970 British Cohort Study (n = 6,575), split 70:30 into training and test sets. Applied directly to the BCS70 test set without optimisation, the UK Biobank-trained model's predictions correlated with observed outcomes at ρ = 0.67; fine-tuning on the BCS70 training set raised this to ρ = 0.72. This is cross-cohort evidence within a second UK cohort, not validation in a different country or health system.

A prioritisation tool, not a diagnostic test

The paper's own framing places RS4DV in primordial prevention: flagging people whose combined biopsychosocial profile warrants earlier assessment, before overt multimorbidity appears. Its value lies in the comparison it enables. A clinician or researcher holding the six routine measures can rank individuals and see that the top decile faces roughly three times the 15-year progression incidence of the bottom decile in the test subset, information that a single risk factor does not supply.

The boundaries are equally specific. The cohort is UK Biobank participants aged 40–70 at recruitment, with a second UK cohort for transfer; performance in other populations is untested here. The score's discrimination is moderate, its incremental C-index gain over conventional covariates plus frailty is about 0.005, and the authors call for prospective studies to determine how it might inform intervention strategies and disparities-focused health planning.

The practical priority is therefore not to treat RS4DV as a verdict but to use it where a cheap, multidomain ranking is useful: deciding who receives further assessment when time, data or assessment burden are constrained. The six-item version is the format the authors position for those settings, and its comparison against its own components is the paper's most decision-relevant result.

Sources used in this answer