OBJECTIVE: Mean intracranial pressure (ICP) is commonly used in the management of patients with intracranial pathologies. However, the shape of the ICP signal over a single cardiac cycle, called ICP pulse waveform, also contains information on the state of the craniospinal space. In this study we aimed to propose an end-to-end approach to classification of ICP waveforms and assess its potential clinical applicability. METHODS: ICP pulse waveforms obtained from long-term ICP recordings of 50 neurointensive care unit (NICU) patients were manually classified into four classes ranging from normal to pathological. An additional class was introduced to simultaneously identify artifacts. Several deep learning models and data representations were evaluated. An independent testing dataset was used to assess the performance of final models. Occurrence of different waveform types was compared with the patients' clinical outcome. RESULTS: Residual Neural Network using 1-D ICP signal as input was identified as the best performing model with accuracy of 93% in the validation and 82% in the testing dataset. Patients with unfavorable outcome exhibited significantly lower incidence of normal waveforms compared to the favorable outcome group even at ICP levels below 20 mm Hg (median [first-third quartile]: 9 [1-36]% vs. 63 [52-88] %, p = 0.002). CONCLUSIONS: Results of this study confirm the possibility of analyzing ICP pulse waveform morphology in long-term recordings of NICU patients. Proposed approach could potentially be used to provide additional information on the state of patients with intracranial pathologies beyond mean ICP.
OBJECTIVE: Mean intracranial pressure (ICP) is commonly used in the management of patients with intracranial pathologies. However, the shape of the ICP signal over a single cardiac cycle, called ICP pulse waveform, also contains information on the state of the craniospinal space. In this study we aimed to propose an end-to-end approach to classification of ICP waveforms and assess its potential clinical applicability. METHODS: ICP pulse waveforms obtained from long-term ICP recordings of 50 neurointensive care unit (NICU) patients were manually classified into four classes ranging from normal to pathological. An additional class was introduced to simultaneously identify artifacts. Several deep learning models and data representations were evaluated. An independent testing dataset was used to assess the performance of final models. Occurrence of different waveform types was compared with the patients' clinical outcome. RESULTS: Residual Neural Network using 1-D ICP signal as input was identified as the best performing model with accuracy of 93% in the validation and 82% in the testing dataset. Patients with unfavorable outcome exhibited significantly lower incidence of normal waveforms compared to the favorable outcome group even at ICP levels below 20 mm Hg (median [first-third quartile]: 9 [1-36]% vs. 63 [52-88] %, p = 0.002). CONCLUSIONS: Results of this study confirm the possibility of analyzing ICP pulse waveform morphology in long-term recordings of NICU patients. Proposed approach could potentially be used to provide additional information on the state of patients with intracranial pathologies beyond mean ICP.