Claire
Najjuuko

Developed and evaluated predictive models to identify adolescents at high risk of suboptimal antiretroviral therapy (ART) adherence in low-resource settings. Applied ensemble methods — Random Forest, Gradient Boosting, and XGBoost — to socio-behavioral and psychosocial data from 702 adolescents living with HIV in Uganda, achieving AUC of 0.71. Models identified key predictors including depression symptoms, stigma experiences, and caregiver support to guide targeted interventions in resource-limited HIV care programs.

Developing a large language model-based system for automated adherence monitoring using routinely collected health record data in HIV clinics. Implementing NLP and LLM pipelines to extract adherence indicators from clinical notes and detect ART non-adherence patterns early, enabling timely clinical intervention in resource-limited settings.

Applied machine learning to longitudinal cohort data from 702 young adults living with HIV in low-resource Ugandan communities to identify key predictors of problematic substance use. Used feature importance analysis and SHAP values to interpret model predictions, surfacing behavioral, psychosocial, and structural risk factors including depression, household instability, and peer influence. Findings inform targeted screening and prevention strategies within HIV care programs.

Designing an AI pipeline that integrates baseline surveys, app engagement metrics, and coach-client text messages from the UMAT-R digital intervention to enable early risk prediction and support personalized recovery care. Developing text-only and multimodal predictive models using TF-IDF/Word2Vec and clinical BERT variants (ClinicalBERT, BioClinicalBERT), combined with survey features. Ongoing work includes transformer-based topic modeling (BERTopic) and self-supervised adaptation of transformer encoders on UMAT-R messages.

Developed and validated a hybrid surveillance system combining rule-based NHSN logic with LLM-driven symptom extraction from clinical notes for catheter-associated urinary tract infection (CAUTI) surveillance. Implemented the Clinical Entity Augmented Retrieval (CLEAR) framework using BIOMed_NER for entity recognition, BioClinicalBERT embeddings for semantic similarity, and GPT-4o for symptom extraction from flagged cases at Barnes-Jewish Hospital (2021–2024). The hybrid system achieved 90.0% sensitivity and 93.5% specificity — outperforming rules-based approaches (80.4% sensitivity) and standalone LLM methods (88.7% sensitivity) — while reducing infection preventionist chart review workload by approximately 66%. Co-first author.

Developed CURA, a clinical uncertainty risk alignment framework for language model-based risk prediction. CURA produces calibrated, uncertainty-aware predictions that enable more reliable and trustworthy clinical decision support.

Building and evaluating conformal prediction-based models for emergency medical services (EMS) that predict intervention protocols, medication types, and quantities from multimodal data combining free-text symptom descriptions and vital signs. Developing calibrated, class-fair prediction sets and selective-classification strategies to provide trustworthy decision support for high-stakes clinical interventions.

Developing and evaluating an uncertainty estimation framework for AI-predicted postoperative complications (acute kidney injury, delirium, deep vein thrombosis, pulmonary embolism, and pneumonia) to support OR-to-ICU handoff decision-making. The framework uses a two-stage conformal prediction approach: in the first stage, a conformal predictor generates preliminary prediction sets with a coverage guarantee; in the second stage, confident singleton rule-outs where the model predicts {Low Risk} alone, are re-evaluated against similar patients in the calibration set. If the rule-out is not well-supported, the prediction is expanded to {Low Risk, High Risk}, signaling uncertainty to the clinician rather than issuing a potentially dangerous confident rule-out. Preliminary results show that this two-stage approach reduces false negative rates across all five complication models. A within-subjects user study will evaluate which uncertainty communication format best supports handoff decisions.

Conducting a benchmark study characterizing uncertainty quantification methods for clinical language models, evaluating their utility and reliability in real-world clinical prediction settings. Comparing approaches including Monte Carlo dropout, ensemble methods, and calibration techniques across multiple clinical tasks.

Analyzed patterns of maternal and child health services utilization and associated socioeconomic disparities across sub-Saharan Africa using machine learning on Demographic and Health Survey data from 352,363 mother-child pairs across 35 countries. Identified critical service utilization patterns linked to child mortality outcomes, providing evidence for targeted policy and health system strengthening.