Work

Machine learning-based risk stratification for neurosyphilis using MRI free-text reports: a 12-year retrospective analysis in China (2012–2024)

Year: 2026
Type: article
Study design: Observational Study
Abstract: Neurosyphilis (NS) has heterogeneous clinical manifestations, while definitive clinical assessment commonly requires invasive cerebrospinal fluid (CSF) testing. Risk stratification using routinely available, non-invasive information could help identify patients who warrant prioritized CSF evaluation. We therefore investigated whether routinely generated brain MRI free-text reports could be systematically transformed into interpretable features for NS risk stratification. In this 12-year retrospective cohort from a national psychiatric specialty center, comprising 4,652 patients with syphilis (319 NS and 4,333 N-NS), we developed and internally validated logistic regression with adaptive LASSO, random forest, and extreme gradient boosting models using structured features extracted from MRI reports, with and without demographic information. All MRI-only models achieved ROC-AUCs above 0.80 in the held-out validation set; for example, the MRI-only adaptive-LASSO model achieved an ROC-AUC of 0.843 (95% CI, 0.786–0.893), with sensitivity of 0.811 and specificity of 0.781. The MRI-plus-demographics adaptive-LASSO model achieved an ROC-AUC of 0.868 (95% CI, 0.818–0.913). Additional analyses indicated that several MRI-report features remained conditionally associated with NS after demographic adjustment. This study developed and internally validated machine-learning models for neurosyphilis risk stratification using structured features extracted from routine MRI free-text reports. These findings suggest that routinely available MRI reports contain clinically relevant information that can be leveraged for NS risk stratification.
Language: English
Field: Medicine
Open Access status: gold
Estimated APC paid: $2,990
OpenAlex ID: W7235922636
FWCI:
Cites:
Cited by:
Related to: 10