gms | German Medical Science

Objective: Spontaneous intracranial hypotension (SIH) is a condition characterized by spinal cerebrospinal fluid (CSF) leakage. As opening pressure on spinal tab is normal in 2/3 of patients, further investigations have identified the resistance to CSF outflow (R_CSF) on lumbar infusion testing as a more sensitive parameter. Most recently R_CSF has been shown to depend on symptom duration. In order to improve the diagnostic yield of lumbar infusion testing we applied classical machine-learning (ML) approaches.

Methods: A total of 182 patients underwent lumbar infusion testing as part of a strict diagnostic protocol to identify spinal CSF leakage as the underlying pathology of orthostatic headache. The training set consisted of 122 patients’ pressure curves. The final evaluation set contained 60 patients’ pressure curves. The feature categories with 1016 features in total were extracted based on spectral analysis and pressure curve morphologies. We followed an approach with classical supervised ML algorithms. Performance metrics were accuracy, sensitivity, specificity and F1 scores.

Results: The R_CSF with a threshold of 5 mmHg/(ml/min) reached only 60% accuracy and 0.51 F1-score for detecting a spinal CSF leak on the final evaluation set, the ML algorithms reached accuracies up to 75% by maintaining F1-scores up to 0.7. The correct diagnosis of a leak rather depends on a set of different feature categories (mainly spectral analysis applications) than on one single CSF parameter. The time from symptom onset until diagnosis did not show to have a major impact using ML approaches.

Conclusion: Analysis of pressure curves can be automated. This may be crucial in diagnosis of complex conditions such as SIH. ML approaches outperformed the conventional lumbar infusion test analysis.