Correcting historical bias in machine learning predictions of healthcare utilization in India
This study investigates how historical disparities in healthcare access influence machine learning (ML) predictions of healthcare utilization among older adults in India through historical bias. We examine prediction differences across disadvantaged populations and quantify their implications for national-level cost projections. Using data from 55,698 respondents in the Longitudinal Ageing Study in India (LASI), we trained two sets of ML models to predict outpatient and inpatient utilization: one on the full population (Model 1) and the other on a subsample classified as having no reported unmet healthcare need (Model 2). Gradient Boosting had the lowest point-estimate RMSE among the benchmark models and was therefore used in the subsequent utilization analyses. To interpret model predictions and identify key drivers of healthcare utilization, we applied SHapley Additive exPlanations (SHAP). We compared model outputs across socioeconomic subgroups and extrapolated predicted utilization to national population estimates using WHO-CHOICE unit costs. Model 2 predictions were higher than Model 1 predictions, particularly among lower-income and caste-identified groups. Overall, outpatient and inpatient predictions from Model 2 were 8.92% (95% CI: 7.96-9.88) and 9.59% (95% CI: 8.59-10.59) higher, respectively. Nationally, this translated to cost differences of I$390.7 (95% CI: 350.6-431.0) million in outpatient care and I$88.4 (95% CI: 79.8-96.9) million in inpatient care. The largest gaps were concentrated in the poorest and most marginalized subgroups. The SHAP analysis suggests that self-rated health (SRH), economic status (MPCE), and chronic conditions are consistently influential in predicting outpatient and inpatient visits, with some shifts in feature importance between models. Differences between models trained on these populations may reflect historical bias arising from unequal access and utilization, together with differences in the covariate distributions of the training populations. These findings underscore the importance of considering data provenance and unequal access when interpreting subgroup differences in healthcare-utilization predictions.
Authors
- Kanya Anindya (ORCID: https://orcid.org/0000-0001-6832-0412)
- Tzu‐Pin Lu (ORCID: https://orcid.org/0000-0003-3697-0386)
- Arokiasamy Perianayagam (ORCID: https://orcid.org/0000-0001-8698-217X)
- Charlotte I. Wang (ORCID: https://orcid.org/0000-0002-6957-1579)
- Rifat Ali Atun (ORCID: https://orcid.org/0000-0002-1531-5983)
- Vincent Cheng-Sheng Li
- John Tayu Lee (ORCID: https://orcid.org/0000-0002-1551-4923)
Institutions
- Harvard University (US)
- National Taiwan University (TW)
- International Institute for Population Sciences (IN)
- University of Gothenburg (SE)
Publication Details
- Journal
- PLOS Digital Health
- Published
- 2026-11-18
- DOI
- https://doi.org/10.1371/journal.pdig.0001815
- Citations
- 1
- Primary Topic
- Artificial Intelligence in Healthcare
- Type
- article
- Field-Weighted Citation Impact
- 2.64
Funders
- Ministry of Education
- Ministry of Education, India