Optimizing Machine Learning Models for Disease Progression Prediction In TB/HIV Coinfection Using Resource-Limited EHR Data Systems in Zambia

Abstract

The integration of artificial intelligence (AI) and machine learning (ML) into electronic health record (EHR) systems offers a transformative Fourth Industrial Revolution (4IR) opportunity for public health decision-making in resource-limited settings. The translation of algorithmic advances into equitable, reliable and actionable clinical tools in low- and middle-income countries (LMICs), however, remains constrained by data quality deficiencies, methodological gaps and governance challenges. This thesis addresses these challenges through three inter related empirical studies conducted in the context of Zambia's national HIV EHR system (SmartCare), drawing on a deduplicated, deterministically linked dataset of 246,053 people living with HIV (PLHIV) across six high-volume public health facilities in Lusaka District. Study One presents a systematic review of 64 peer-reviewed studies published between 2018 and 2025 on ML-based predictive analytics using EHR data, with LMICs as the primary focus. Conducted in accordance with PRISMA 2020 guidelines and using a PROBAST-AI-adapted risk-of-bias instrument, the review establishes that only 18.8% of studies were conducted in LMIC settings despite these settings carrying the greatest burden of preventable disease. External validation was reported in only 7.8% of studies, calibration in 6.2%, and explainability in 8.3% of LMIC-focused studies compared with 36.4% in high-income settings. Four cross-cutting evaluation, contextual, ethical and deployment gaps are identified and used to scope the empirical chapters. Study Two develops and evaluates a leakage-safe, multi-outcome, externally validated and explainable ML framework predicting three 12-month binary outcomes among PLHIV: recorded TB treatment (TB12m), programme-recorded interruption in treatment (IIT12m) and recorded unsuppressed viral load (UVL12m). A stacked ensemble of logistic regression, random forest and XGBoost with a logistic regression meta-learner achieved AUC-ROC of 0.707 for TB12m, 0.839 for IIT12m and 0.778 for UVL12m on a held-out facility test set (n = 74,506). F1 optimised decision thresholds support operational protocols ranging from high-sensitivity screening (TB12m, threshold 0.257, sensitivity 92.3%) to highly targeted intensive intervention (IIT12m and UVL12m, thresholds near 0.9). SHAP-based local interpretability shows that education level is the dominant predictor of UVL12m and a leading predictor of IIT12m, raising critical questions about whether the signal reflects genuine clinical risk or facility-level data recording patterns. Facility-holdout validation confirms cross-site generalisability under deployment-realistic conditions. Study Three conducts a systematic eight-dimension data quality assessment (DQA) of the SmartCare EHR system using the AI pipeline as a diagnostic lens. Education-level completeness ranged from 1.2% to 7.9% across all six facilities; engagement feature coverage was only 12.3%; and 54.3% of patients with engagement data had implausible appointment-lateness values, the highest being 45,459 days (approximately 124 years). The composite Data Quality Index (DQI) ranged from 53.5% to 58.3%, all below the 70% minimum threshold proposed in the thesis for iv fully trustworthy AI training. Eight minimum data quality requirements and a reproducible DQA toolkit are proposed, with quantitative remediation targets at facility and programme levels. Complementing these empirical studies, the thesis develops a 4IR health transformation framework situating AI adoption within Zambia's broader digital health ecosystem; constructs an economic value creation model demonstrating positive cumulative cost-benefit by Year 3 of deployment under base-case assumptions; and proposes an institutional governance architecture together with a five-phase AI adoption roadmap (2025-2030) aligned with PEPFAR, Global Fund and Ministry of Health programmatic cycles. A proposed ICT integration architecture linking SmartCare to an HL7 FHIR R4-compliant ML prediction engine, DHIS2, DISA and a clinical decision support dashboard provides the engineering blueprint for transitioning from research-grade pipelines to production-ready clinical deployment at national scale. Two formal dissemination meetings were convened during the research: with the Ministry of Health Provincial Health Office Monitoring and Evaluation Data Management teams (representatives from Lusaka, Southern, Western, North-Western and Eastern Provinces); with the AI Think-Tank Team of experts at the Centre for Infectious Disease Research in Zambia (CIDRZ). Feedback across these forums was uniformly positive and is integrated into the thesis recommendations. Together, these contributions show that trustworthy AI in resource-limited public health requires as much investment in data governance, economic justification, ICT architecture and institutional capacity as in algorithmic development. The thesis offers an integrated framework, evidence base and operational roadmap for responsible AI deployment in Zambia's HIV programme, with direct relevance to comparable settings across sub-Saharan Africa.