Machine learning-based models using preoperative imaging and clinical data to predict intraoperative rebleeding and functional recovery in intracerebral hemorrhage patients: a multicenter study.
Authors
Affiliations (8)
Affiliations (8)
- Department of Neurosurgery, Traditional Chinese Medicine Hospital of Xinjiang Uygur Autonomous Region, Urumqi, China.
- Faculty of Medicine, Taylor's University, Selangor, Malaysia.
- Department of Neurosurgery, Traditional Chinese Medicine Hospital of Xinjiang Uygur Autonomous Region, Urumqi, China - [email protected].
- Department of Interventional Medicine, Tongji Hospital Affiliated to Tongji University, Shanghai, China.
- Intervention Center, The Sixth Affiliated Hospital of Xinjiang Medical University, Urumqi, China.
- Shanghai Traditional Chinese Medicine Hospital, Shanghai, China.
- Chongqing Hospital of Jiangsu Provincial Hospital of Traditional Chinese Medicine, Chongqing, China.
- The Fifth Affiliated Hospital of Xinjiang Medical University, Urumqi, China.
Abstract
Spontaneous intracerebral hemorrhage (ICH) is associated with high mortality and disability. This study aimed to develop and validate machine learning models based on preoperative multimodal data to predict the risk of intraoperative rebleeding (IOR) and postoperative functional recovery in ICH patients, thereby providing support for precision treatment decisions. This multicenter prospective cohort study included 498 primary ICH patients who underwent surgical treatment at three tertiary hospitals in China between January 2018 and December 2023. Preoperative clinical characteristics, laboratory tests, and imaging parameters were collected. The primary endpoint was IOR (defined as obvious arterial bleeding requiring additional hemostatic measures during surgery, or residual/new hematoma increased by ≥5 mL on CT within 24 hours after surgery compared to preoperative assessment). The secondary endpoint was good functional outcome (modified Rankin Scale [mRS] ≤2) at 6 months postoperatively. Machine learning algorithms were used to construct prediction models, which were evaluated through temporal validation and external multicenter validation cohorts. Among the included patients, the incidence of IOR was 14.7% (73/498), and the rate of good outcome at 6 months was 42.6% (212/498). After rigorous cross-validation and hyperparameter optimization, the deep neural network model performed best in predicting IOR (AUC 0.892, 95%CI: 0.851-0.933), with accuracy of 83.7%, sensitivity of 84.6%, and specificity of 83.5%. The XGBoost model was optimal for predicting functional recovery (AUC 0.875, 95%CI: 0.831-0.919), with accuracy of 81.9%, sensitivity of 82.1%, and specificity of 81.8%. In the external validation cohort (N.=156), the two models achieved AUCs of 0.842 (95%CI: 0.768-0.916) and 0.831 (95%CI: 0.758-0.904), respectively. Multivariate analysis showed that hematoma volume (OR=1.35, 95%CI: 1.21-1.51, per 10 mL increase), admission INR (OR=3.17, 95%CI: 2.04-4.93, >1.4 vs. ≤1.4), midline shift (OR=1.28, 95%CI: 1.13-1.44, per 1 mm increase), and preoperative platelet count (OR=0.83, 95%CI: 0.76-0.91, per decrease of 20×10<sup>9</sup>/L) were independent risk factors for IOR. Decision curve analysis demonstrated that machine learning models provided higher net clinical benefit than traditional ICH and FUNC scores. Machine learning prediction models validated across multiple centers can accurately assess the risk of IOR and postoperative functional recovery in ICH patients, thereby guiding individualized treatment decisions and potentially improving patient outcomes.