Predicting call abandonment in a mental health helpline: A calibrated machine-learning approach

Mental health helplines regularly lose callers before they reach a counsellor, and these abandoned calls represent unmet demand. Real-time methods to predict abandonment remain scarce. As part of a KTP project developing a digital analytics platform for mental health and wellbeing services, we built a machine-learning model that estimates abandonment risk while callers are in queue and flags high-risk calls to support real-time operational decisions. The study used 141,952 calls from 2023 to 2025, with an abandonment prevalence of 1.72%. We used a strict temporal design: training on 2023 data with tuning and threshold selection on 2024 data, and held-out evaluation on 2025 data. Outlier caps, aggregate features, and preprocessing were fitted on training data only to prevent leakage. A set of 15 features captured call routing, wait time, and automated menu engagement. Gradient boosting was selected on PR-AUC, with a feature-group ablation confirming the chosen features. Probabilities were calibrated with isotonic regression using time-aware cross-validation folds. The calibrated model achieved a test PR-AUC of 0.21 and ROC-AUC of 0.93. An operational threshold of 0.85, selected from the 0.80 to 0.85 range with priority on recall, flagged 2,383 calls with an abandonment rate of 24.3% (around 14 times the base rate). Operationally, this supports resource allocation during periods of elevated risk, while the ethics of any caller-level response require separate consideration.