Prognostic value of early intracranial pressure curve following surgical decompression: A machine learning analysis

Intracranial pressure (ICP) control is a critical determinant of successful therapy following decompressive craniectomy (DC). While traditional assessments have focused on peak or mean ICP values, the prognostic relevance of continuously quantified ICP burden using the area under the curve (ICP-AUC) remains underexplored. This study aimed to evaluate the prognostic value of early postoperative ICP dynamics within the first 72 h using ICP-AUC based analysis and to assess predictive performance through machine learning models.

Seventeen patients who underwent DC were included in this retrospective analysis. ICP was continuously recorded over a 72-h period after DC was performed. We calculated normalized ICP-AUC, maximum ICP values, and the ICP-AUC above thresholds was calculated separately based on baseline ICP values of 15 mmHg and 20 mmHg, respectively. Neurological outcomes were assessed using the modified Rankin Scale (mRS) at discharge. A poor outcome was defined as mRS = 6 (death).

A normalized ICP-AUC ≥0.5 was significantly associated with poor outcomes at discharge (Spearman r = 0.50, p = 0.04), and all patients with ICP-AUC ≥0.6 deceased. Furthermore, ICP levels exceeding 20 mmHg for more than 30% of the time were significantly associated with mRS = 6 (p = 0.02). Additional threshold analyses confirmed significance of ICP-AUC based on threshold of >15 mmHg (p = 0.047) >20 mmHg (p = 0.03). ROC analysis revealed an AUC of 0.89 for the 20 mmHg threshold, indicating high predictive accuracy. The Random Forest model achieved an AUC of 0.89, with a precision of 1.00 and an F1-score of 0.80. In contrast, the XGBoost model showed lower predictive values across all metrics (0.67). On the other hand, models based on a 15 mmHg ICP baseline demonstrated limited prognostic validity.

Early cumulative ICP burden, assessed through ICP-AUC, might provide a reliable prognostic marker for neurological outcomes following DC. Machine learning enhances predictive accuracy and offers a promising approach for clinical decision support in neurocritical care.