Background Cerebrospinal fluid (CSF) culture is essential for diagnosing neuroinfectious diseases such as meningitis, where clinical symptoms alone cannot reliably confirm the diagnosis or differentiate bacterial from non-bacterial causes. However, traditional CSF cultures are time-consuming and often lack sensitivity, particularly after antibiotic treatment. This study applies machine learning to expedite diagnosis and enhance the accuracy of outcome predictions for patients with suspected neuroinfectious diseases.

Methods Training and validation datasets were derived from the MIMIC-III and MIMIC-IV databases, respectively, focusing on patients undergoing CSF bacterial culture. CNN, FCN, and RF classifiers were trained using integrated test results and clinical data to predict CSF culture outcomes. CSF culture result labels were redefined based on clinical guidelines for bacterial meningitis diagnosis to address false negatives caused by prior antibiotic treatment. Unstructured text data were encoded using BioBERT embeddings, combined with structured laboratory data, and processed through CNN, FCN, or LSTM layers to develop multimodal neural networks for predicting clinical outcomes. Class imbalance was mitigated, and text data sources were optimized for interchangeability.

Results For CSF culture diagnostic predictions, 9261 cases were analyzed, including 245 initial culture-positive and 9016 culture-negative results. Label correction increased the positive cases to 404, with 8857 negatives. The FCN model showed the best performance post-label correction, achieving an AUC of 0.891, accuracy of 0.867, sensitivity of 0.833, and specificity of 0.868. In survival prediction analysis (5795 cases: 5500 survivors, 295 in-hospital deaths), the optimized multimodal neural network with CNN layers achieved an AUC of 0.933, accuracy of 0.850, sensitivity of 0.847, and specificity of 0.850. Models trained without text data showed diminished performance. Independent validation datasets confirmed the model’s robustness, yielding comparable results.

Conclusions The multimodal machine learning models demonstrated strong efficacy in predicting CSF culture outcomes, supporting the diagnosis of bacterial meningitis, and in forecasting ICU survival outcomes. Integrating these models into a clinical decision support system could empower ICU clinicians with timely, data-driven insights, enabling culture-independent diagnostics and enhancing patient management.

The authors have declared no competing interest.

This work was supported by National Natural Science Foundation of China (81672627; 82071863).

I confirm all relevant ethical guidelines have been followed, and any necessary IRB and/or ethics committee approvals have been obtained.

Yes

The details of the IRB/oversight body that provided approval or exemption for the research described are given below:

The research utilized the MIMIC database (physionet.org/content/mimiciv/3.1/), which was initially approved for use by the Institutional Review Boards of Beth Israel Deaconess Medical Center and the Massachusetts Institute of Technology. Since MIMIC is a publicly available, de-identified dataset, individual patient consent was not required. All analyses and research activities strictly adhered to the MIMIC data access agreement and data usage guidelines.

I confirm that all necessary patient/participant consent has been obtained and the appropriate institutional forms have been archived, and that any patient/participant/sample identifiers included were not known to anyone (e.g., hospital staff, patients or participants themselves) outside the research group so cannot be used to identify individuals.

Yes

I understand that all clinical trials and any other prospective interventional studies must be registered with an ICMJE-approved registry, such as ClinicalTrials.gov. I confirm that any such study reported in the manuscript has been registered and the trial registration ID is provided (note: if posting a prospective study registered retrospectively, please provide a statement in the trial ID field explaining why the study was not registered in advance).

Yes

I have followed all appropriate research reporting guidelines, such as any relevant EQUATOR Network research reporting checklist(s) and other pertinent material, if applicable.

Yes

The data utilized in this study are derived from the MIMIC-III (https://doi.org/10.13026/C2XW26) and MIMIC-IV (https://doi.org/10.13026/a3wn-hq05) critical care databases, which are accessible to researchers who fulfill the credentialing requirements set by the data custodians. Due to these restrictions, the raw data are not publicly available. However, upon obtaining the necessary permissions from the MIMIC data providers, the data can be accessed for replication or further study. The complete set of code for implementing the machine learning models, including those predicting CSF culture results and ICU outcomes, is openly available in our GitHub repository: https://github.com/AaronChen007/CNS_Model_Code.