# Development and Validation of a Novel Risk Calculator to Predict Sub-optimal HIV Outcomes Among Pregnant and Postpartum Women with HIV in Kenya

## TL;DR

This study created a risk calculator to predict poor HIV outcomes in pregnant and postpartum women in Kenya, helping identify those needing targeted support.

## Contribution

A novel risk calculator was developed and validated to predict disengagement from care and treatment failure in pregnant and postpartum women with HIV.

## Key findings

- The risk calculator includes demographic, clinical, and psychosocial variables to predict outcomes.
- The model for women with known HIV diagnosis showed strong discrimination (AUROC 0.843) and moderate calibration.
- The calculator can stratify risk into low, moderate, and high categories for targeted interventions.

## Abstract

No tool currently exists to predict the cumulative risk of suboptimal clinical outcomes among pregnant and postpartum women with HIV (PPWH). This study sought to develop and validate a parsimonious risk calculator capable of predicting disengagement from care and HIV treatment failure among PPWH. We created the risk calculator using data from 1,331 PPWH from Southwestern Kenya (Homabay, Migori, and Kisumu Counties) in the Mother Infant Visit Adherence and Treatment Engagement Trial. Least absolute shrinkage and selection operator logistic regression retained the most predictive variables from 16 candidate factors to estimate the probability of treatment failure or disengagement from care. Three risk quintiles were calculated. We assessed external validation with an independent dataset (Opt4Mamas; N = 820). Cross-validated area under the curve of receiver operating characteristic (AUROC) and calibration measures assessed model performance. Two unique risk calculators were created - one for PPWH with known HIV diagnosis prior to pregnancy and one for PPWH with new HIV diagnoses. The combined outcome of care disengagement or treatment failure occurred in 43% of PPWH with known diagnosis and 40% with new diagnosis in the development dataset; and 37% with known diagnosis and 13% with new diagnosis in the validation dataset. The calculators included demographic (age, parity, marital status), clinical (virological failure, missed visits, regimen line, gestation age), and psychosocial variables (intimate partner violence, stigma, depression, partner support, disclosure, food insecurity). The model for PPWH with known diagnosis demonstrated better calibration and discrimination (AUROC 0.843 [95% CI 0.805, 0.866]) than the model for PPWH with a new HIV diagnosis (AUROC 0.463 [95% CI 0.347, 0.597]). Mean predicted risk probabilities among PPWH with known HIV diagnosis were: low (6%), moderate (56%), and high (70%). Mean predicted risk probabilities among those with a new HIV diagnosis were: low (31%), moderate (48%), and high (65%). The novel risk calculator for PPWH with a known HIV diagnosis has the potential to identify those who are at risk of sub-optimal HIV treatment and care outcomes for targeted interventions to prevent treatment failure and loss to follow-up.

The online version contains supplementary material available at 10.1007/s10461-025-04814-8.