Comparative Analysis of Machine Learning Models for ICH Prognostication Using Diverse Datasets

Multiple prognostic scores have been developed to predict morbidity and mortality in patients with spontaneous intracerebral hemorrhage(sICH). These scoring models were traditionally based on statistical methods involving a limited set of variables. Machine learning (ML) has enabled the development of prognostic models for spontaneous ICH which can leverage much more data. We trained ML models on two distinct datasets: (1) Qatar dataset only, and (2) a combined dataset consisting of Qatar and ATACH datasets. Model validation was conducted separately on the Qatar and ATACH test sets, providing insights into model performance within and across study populations. By incorporating inpatient variables into model development, we leveraged more information. For 90‐day mortality prediction, the RF model trained on the combined dataset demonstrated superior AUC on the ATACH test set (0.945) compared to the model trained only on Qatari data (0.896). Similarly, XBG showed significant improvement in AUC from 0.889 to 0.950 when trained on the combined dataset. RF model gave the best balance of high AUC + low Brier across both test sets (ATACH and Qatar). For functional outcome prediction, a similar trend was observed. The RF model trained on the combined dataset had improved prediction (AUC 0.878 vs 0.841), and so did the XGB model trained on the combined dataset (AUC 0.885 vs 0.821). For 90‐day functional outcomes, XGBoost (trained on combined dataset) had the best discriminative power and the best calibration. Amongst the different machine learning models tested, Random Forest (RF) demonstrated the most balanced performance, achieving high metrics across both mortality and functional outcomes. When comparing the models trained exclusively on Qatari data versus those trained on the combined dataset (Qatar + ATACH), the combined dataset models generally showed improved generalization. This is an interesting observation given our datasets were very diverse with patients from multiple ethnicities, diverse backgrounds, and having different risk profiles. It also indicates that models that are trained on a more diverse patient population capture broader underlying clinical patterns and ehance their applicability to diverse patient populations. Our study design mirrors a realworld deployment scenario in which a model developed at a single centre is transported to external cohorts, while still preventing any information leakage from test data. Future studies can involve federated learning approaches to big data.