Algorithmic fairness and transportability of machine learning models predicting cancer screening uptake in India: a nationally representative study
Year: 2026
Type: preprint
Abstract: Background Machine learning models built from national surveys are increasingly proposed to identify people who are unlikely to be screened for cancer, so that outreach can be targeted. Whether such models capture individual behaviour or merely reproduce where programmes operate, whether they transfer to new settings, and whether they treat population groups equitably is rarely examined, particularly in low- and middle-income countries. Methods We used India's fifth National Family Health Survey (2019-21) to predict ever having had cervical, breast and oral cancer screening among 357353 women and oral screening among 45739 men aged 30–49 years. Logistic regression and XGBoost models were trained with individual predictors, area predictors (state and out-of-fold district screening rate) or both. We assessed discrimination, calibration, survey-weighted performance, SHAP explanations and decision curves; transportability by leave-one-state-out validation with random-effects pooling; fairness across sex, wealth, caste, residence and religion; two bias-mitigation strategies; and the effect of resampling before the train-test split. Results Individual predictors alone achieved AUCs of 0.58–0.69, whereas area predictors alone achieved 0.63–0.82 and combined models 0.67–0.82, with good calibration. Area features accounted for 59–69% of SHAP importance in women. When a state was held out, individual-level models performed at chance (pooled AUC 0.51–0.54) and observed screening ranged from 0.17 to 15 times the predicted level. When the top 10% of predicted risk was flagged, sensitivity among the poorest women was 0.18–0.38 compared with 0.57–0.68 among the richest two quintiles, giving equal opportunity differences of 0.30–0.45 across wealth and 0.31–0.46 across caste or tribe groups. Group-specific thresholds reduced the wealth disparity to 0.09–0.10 without loss of AUC, and reweighing reduced it to 0.08 for cervical screening. Applying synthetic minority oversampling before the train-test split inflated the AUC from 0.64 to 0.97. Conclusions Survey-based screening-uptake models predict where screening programmes operate rather than who takes up screening, do not transfer across states, and allocate outreach unequally unless fairness is explicitly addressed. Such models should not be used to target individuals, and published high accuracies should be interpreted with caution where resampling precedes data splitting.
Source: Research Square
Language: English
Subfield: Oncology
Field: Medicine
Domain: Health Sciences
Open Access status: green
OpenAlex ID: W7225700792