JMIR Publications

ABSTRACT

Background:

Asthma attacks remain a significant burden on primary care systems, yet predicting individual risk remains challenging. While machine learning shows promise, most existing models are derived from secondary care data or lack robust validation and assessment of clinical utility.

Objective:

This study aims to develop and validate machine learning models for predicting 12-month asthma attack risk using UK-wide primary care data, adhering to TRIPOD-AI and RECORD reporting guidelines.

Methods:

We conducted a retrospective cohort study using the Optimum Patient Care Research Database (OPCRD), comprising more than 600,000 patients with active asthma across England, Scotland, and Wales. Data were split into a training set (England) and a geographical validation/test set (Scotland and Wales). We compared five algorithms: Logistic Regression (LR) as the benchmark model, Decision Tree (DT), Random Forest (RF), Extreme Gradient Boosting (XGBoost), and neural networks (NN) (a combination of fully connected network [FCN] and long short-term memory [LSTM]). We assessed model performance using the Area Under the Receiver Operating Characteristic (ROC) curve (AUC), calibration plots, and Decision Curve Analysis (DCA) to quantify net benefit.

Results:

XGBoost emerged as the top-performing model on both the validation (AUC = 0.816) and test (AUC = 0.788) sets. It achieved 5% higher sensitivity than LR (95% vs 90%) at 62% specificity. At the same specificity (62%), XGBoost identified about five more patients per 100 who had an asthma attack within 12 months than LR, indicating a higher net benefit. Additionally, all other models outperformed LR on the test set (LR AUC: 0.754; DT: 0.770; RF: 0.783; NN: 0.786). Model calibration confirmed reliable probability estimates, and the model interpretation highlighted key predictors, including prior asthma attacks and medication use.

Conclusions:

These findings indicate that advanced ML approaches substantially improve asthma risk prediction over a conventional regression model. Future work should focus on prospective evaluation and implementation science to determine how to integrate these models into primary care workflows to improve preventive intervention targeting and reduce asthma attacks.