Purpose Intracranial hemorrhage (ICH) requires urgent treatment, and accurate and timely diagnosis is essential for improving outcomes. This pivotal clinical trial aimed to validate a deep learning algorithm for ICH detection and assess its clinical utility through a reader performance test. Methods Retrospective CT scans from patients with and without ICH were collected from a tertiary hospital. Two experts evaluated all scans, with a third expert reviewing disagreements for the final diagnosis. We analyzed the performance of the deep learning algorithm, JLK-ICH, for all cases and ICH subtypes. Additional external validation was performed using a multi-ethnic U.S. dataset. A reader performance study included six non-expert readers who evaluated 800 CT scans, with and without JLK-ICH assistance, following a washout period. ICH presence and five-point scale confidence level for decisions were rated. Results A total of 1,370 CT scans were evaluated. The deep learning model showed 98.7% sensitivity (95% confidence interval [CI] 97.8-99.3%), 88.5% specificity (95% CI, 83.6-92.3%), and an area under the receiver operating characteristic curve (AUROC) of 0.936 (95% CI, 0.915–0.957). The model maintained high accuracy across all ICH subtypes, and additional external validation confirmed these results. In the reader performance study, AUROC with JLK-ICH assistance (0.967 [0.953–0.981]) surpassed that without assistance (0.953 [0.938–0.957]; P = 0.009). JLK-ICH particularly improved performance when readers were highly uncertain. Conclusion The JLK-ICH algorithm demonstrated high accuracy in detecting all ICH subtypes. Non-expert readers significantly improved diagnostic accuracy for brain CT scans with deep learning assistance.
Although large language models (LLMs) have demonstrated impressive reasoning capabilities across general domains, their effectiveness in real-world clinical practice remains limited. This is likely due to their insufficient exposure to real-world clinical data during training, as such data is typically not included due to privacy concerns. To address this, we propose enhancing the clinical reasoning capabilities of LLMs by leveraging real-world clinical data. We constructed reasoning-intensive questions from a nationwide sepsis registry and fine-tuned Phi-4 on these questions using reinforcement learning, resulting in C-Reason. C-Reason exhibited strong clinical reasoning capabilities on the in-domain test set, as evidenced by both quantitative metrics and expert evaluations. Furthermore, its enhanced reasoning capabilities generalized to a sepsis dataset involving different tasks and patient cohorts, an open-ended consultations on antibiotics use task, and other diseases. Future research should focus on training LLMs with large-scale, multi-disease clinical datasets to develop more powerful, general-purpose clinical reasoning models.