Explainable Prediction Of Digital Brain Function Screen Score From Self-reported Lifestyle And Clinical Context

Jermyn See¹, See Ann Soo¹, Nav Vij¹, Prem Pillay²

¹Neurowyzr Pte Ltd, Singapore, ²Singapore Brain Spine Nerves Center, Singapore

SUMMARY

A 500x bootstrap-bagged Explainable Boosting Machine (EBM) trained on N=4,345 routine screenings (Singapore & Malaysia) predicts Digital Brain Function Screen (DBFS) cognitive score from self-reported lifestyle and demographic context (R²=0.554, r=0.744, RMSE=21.6). Cognitively stimulating activity (~14 pts) and education (~7 pts) show the largest feature-response spans, with transparent per-patient explanations supporting structured clinician-led lifestyle counselling.

INTRODUCTION

Background

Lifestyle factors are modifiable determinants of cognitive decline.

Most screening pathways lack a practical way to translate self-report into individualised guidance alongside the test result.

Black-box models don't fit clinic workflows; clinicians need to see why a score is what it is before acting on it.

Objective

Build an interpretable model paired with DBFS that estimates expected screening performance from routine self-report and produces transparent, patient-specific explanations for clinician-led follow-up.

About DBFS

15-minute digital cognitive assessment: memory, executive function, attention, validated for use in routine health screening.

Returns an overall DBFS score that is a readout of cognitive health and indicator of MCI risk.

Includes a self-reported lifestyle questionnaire (Table 1).

METHODS

Data

N = 4345, routine health screenings, Singapore & Malaysia

Exclusion criteria: individuals with diagnosed dementia or major neurologic/psychiatric illness.

Model: Explainable Boosting Machine (EBM)

Glassbox additive model: each feature's contribution is directly readable as a term-contribution shape function.

500× bootstrap-bagging; term contributions averaged across resamples for stability; ±1 SD bands shown per term.

Evaluation

Held-out 80/20 split (single constrained EBM) as the generalisation estimate (Fig 1); deployed model is the full 500× bag.

RESULTS

Held-out performance: R²=0.554, RMSE=21.6, r=0.744 (n=869); ceiling compression above ~130 pts

Largest modifiable effects: cognitively stimulating activity (~14.4 pts) and education (~6.9 pts); wide stability bands on rare categories (physical activity, sleep) reflect data sparsity, not instability (Fig. 2)

Per-patient explanations are fully auditable: baseline + signed feature contributions sum to the predicted score (Fig. 3)

CONCLUSIONS

Bagged EBM predicts DBFS performance from routine self-report with transparent, per-patient explanations, no extra data collection needed.

Largest modifiable effects (cognitive activity, education) are consistent with cognitive reserve; appropriate for contextualising scores, not causal targets.

Supports structured lifestyle counselling as a direct add-on to DBFS screening

Limitations: internal evaluation only (external/temporal validation pending); cross-sectional. Figures reflect updated model (N=4,345); metrics differ slightly from submitted abstract.

REFERENCES 

1. Livingston G, et al. Lancet. 2024;404:572–628. 
2. Lou Y, Caruana R, Gehrke J. Proc. ACM SIGKDD. 2012:150–158. 
3. Nori H, et al. InterpretML. arXiv:1909.09223. 2019. 
4. Liu Y, et al. Front Aging Neurosci. 2024;16:1358992. 
5. Chew EY, et al. Bioinformation. 2023;19(5):522–524.

For further inquiries about the study, please contact Jermyn See at jermyn.see@neurowyzr.com.