Feasibility of the optimal cerebral perfusion pressure value identification without a delay that is too long

Abstract

Optimal cerebral perfusion pressure (CPPopt)-targeted treatment of traumatic brain injury (TBI) patients requires 2-8 h multi-modal monitoring data accumulation to identify CPPopt value for individual patient. Minimizing the time required for monitoring data accumulation is needed to improve the efficacy of CPPopt-targeted therapy. A retrospective analysis of multimodal physiological monitoring data from 87 severe TBI patients was performed by separately representing cerebrovascular autoregulation (CA) indices in relation to CPP, arterial blood pressure (ABP), and intracranial pressure (ICP) to improve the existing CPPopt identification algorithms. Machine learning (ML)-based algorithms were developed for automatic identification of informative data segments that were used for reliable CPPopt, ABPopt, ICPopt and the lower/upper limits of CA (LLCA/ULCA) identification. The reference datasets of the informative data segments and, artifact-distorted segments, and the datasets of different clinical situations were used for training the ML-based algorithms, allowing us to choose the appropriate individualized CPP-, ABP- or ICP-guided management for 79% of the full monitoring time for the studied population. The developed ML-based algorithms allow us to recognize informative physiological ABP/ICP variations within 24 min intervals with an accuracy up to 79% (compared to the initial accuracy of 74%) and use these segments for timely optimal value identification or CA limits determination in CPP, ABP or ICP data. Prospective clinical studies are needed to prove the efficiency of the developed algorithms.

Figure 1

Problematics of CPPopt value identification and research idea. Machine learning (ML)-based algorithms are applied to recognize the informative physiological ABP/ICP variations within short (20–30 min) intervals and use these data episodes for the timely identification of optimal value (or lower/upper CA limits) in CPP, ABP or ICP data. CPPopt(t) value is identified according to recognized “informative” data episodes (marked in red) as a minimum point of “U-shape” approximation in PRx = f(CPP) graph.

Figure 2

Software tool for ICU multimodal monitoring data analysis and CPPopt identification. The contribution of ABP and ICP color-marked episodes (i.e., green, cyan, blue, pink and red) to the formation of quadratic approximation curves can be analyzed by including or removing them from the PRx = f(CPP), PRx = f(ABP), and PRx = f(ICP) graphs. Color-marked segments are selected from software menu by assigning segments into “artifact-free”/”artifact-disturbed” or “informative”/”noninformative” classes which are required to train and test the ML algorithms. The example in (A) represents the case when CPPopt and ABPopt together with the upper limits of autoregulation (ULCA) are identified, thus demonstrating that CPPopt-targeted treatment can be applied for the personalized management to control intact CA. The example in (B) shows that CPPopt identification can be disturbed due to an ICP increase. However, the detected ULCA for ICP shows that ICP increase above 26 mmHg leads to impaired CA. Here ND is “not detected”.

Figure 3

Workflow diagram of multimodal monitoring data analysis using ML models for CPPopt, ABPopt, LLCA, and ULCA identification.

Figure 4

Examples of CPPopt identification according to informative ABP and ICP episodes (marked in red). The PRx = f(CPP) graph demonstrated the contribution to informative data points (pink) and noninformative data points (gray) to the formation of the U-shape approximation and CPPopt identification results. The red curve in the PRx = f(CPP) graph represents the U-shape approximation when ML is applied, and the black curve is the approximation when ML is not applied. These results show that short informative episodes (20–30 min) with informative physiological ABP and ICP data changes are sufficient to obtain the U-shape and identify the CPPopt value. (a)—Without ML (black color) CPPopt = 84 mmHg, LLCA is not detected and ULCA = 102 mmHg. With the ML algorithm (red color) CPPopt = 78 mmHg, LLCA not detected and ULCA = 95 mmHg. (b)—Without ML (black color) CPPopt is not detected, LLCA = 73 mmHg and ULCA is not detected. With the ML algorithm (red color) CPPopt = 82 mmHg, LLCA = 69 mmHg and ULCA is not detected. (c) Without ML (black color) CPPopt = 100 mmHg, LLCA = 89 mmHg and ULCA = 112 mmHg. With the ML algorithm (red color) CPPopt = 100 mmHg, LLCA = 88 mmHg and ULCA = 112 mmHg. (d) Without ML (black color) CPPopt = 84 mmHg, LLCA = 74 mmHg and ULCA = 94 mmHg. With the ML algorithm (red color) CPPopt = 88 mmHg, LLCA = 81 mmHg and ULCA = 95 mmHg. Here ND is “not detected”.

Figure 5

Representation of the diagnostic ability of the developed ML model to recognize CPP-, ABP- or ICP-guided management, as demonstrated by the receiver operating characteristic (ROC) curves. ROC characteristics for recognition CPP-guided management: sensitivity 68%, specificity 66%, AUC 0.74; ABP-guided management: sensitivity 51%, specificity 81%, AUC 0.74, ICP-guided management: sensitivity 31%, specificity 89%, AUC 0.75; recognition of critical status: sensitivity 87%, specificity 97%, AUC 0.98; recognition of intact CA status: sensitivity 77%, specificity 94%, AUC 0.95.