Evaluating the effectiveness of non-invasive intracranial pressure monitoring via near-infrared photoplethysmography using classical machine learning methods

This study investigates the feasibility of utilising photoplethysmography signals to estimate continuous intracranial pressure (ICP) values in patients with traumatic brain injury. A clinical dataset was compiled, comprising synchronised data from a non-invasive optical sensor and an invasive gold standard ICP monitor from 27 patients. Two datasets, derived from short and long-distance NIRS, were generated from this data. For each dataset, 141 features were extracted for every one-minute window of non-invasive data. A total of 5 regression models were assessed. The study aimed to evaluate the models' performance for the continuous, noninvasive monitoring of ICP using a leave-one-patient out cross validation approach. The 5 models were trained on both the long and short distance NIRS data. The lowest mean absolute error (MAE) and root mean squared error (RMSE) were obtained using features derived from long-distance NIRS. A Random Forest (RF) model achieved the lowest MAE and RMSE of 5.030 and 4.067 mmHg respectively. The RF exhibited wide limits of agreement with the reference method. This was reflected in the 95% Bland-Altman limits of agreement, ranging from 8.782 to -8.487 mmHg.