Analysis of neurocognitive status after cardiopulmonary bypass surgery using mathematical methods

Cardiac surgery, particularly coronary artery bypass grafting (CABG), represents one of the most critical interventions in modern cardiovascular medicine. Cardiopulmonary bypass (CPB) is an essential technique in open-heart surgery, where a heart-lung machine temporarily assumes the functions of the heart and lungs, allowing surgeons to operate on a still heart in a bloodless field. While CPB has revolutionized cardiac surgery and enabled complex procedures that save countless lives, it is not without complications. One of the most significant concerns following cardiac surgery is the development of postoperative neurocognitive dysfunction (POCD), which can substantially impact patients' quality of life and recovery. Postoperative neurocognitive dysfunction encompasses alterations in cognitive functions including memory, attention, executive functions, language skills, and visuospatial abilities. Despite advances in surgical techniques, anesthesia protocols, and medical device technologies, POCD remains prevalent following cardiac surgery, with reported incidence rates ranging from 24% to 79% in various studies. The etiology of POCD after CABG surgery is multifactorial and not fully understood, though intraoperative hypotension and multiple emboli are considered two primary mechanisms leading to brain injury. The relationship between hemodynamic parameters during CPB and subsequent neurocognitive outcomes has been the subject of considerable research, yet significant questions remain unanswered. Optimal mean arterial pressure (MAP) during CPB to ensure adequate tissue perfusion remains controversial, with some researchers advocating for lower targets (50-60 mmHg) while others recommend higher values (70-80 mmHg). Multiple factors can influence MAP during surgery, including flow rate, blood viscosity (affected by temperature and hematocrit), depth of anesthesia, anesthetic agents used, and perioperative inflammation. Understanding these complex interactions and their impact on cognitive outcomes represents a critical challenge in cardiac surgery. The primary aim of this study is to comprehensively analyze the effects of detailed hemodynamic data recorded during CABG surgeries on postoperative neurocognitive dysfunction. Unlike previous studies that primarily employed basic statistical analyses, this research leverages advanced machine learning algorithms and artificial neural networks to examine the complex relationships between hemodynamic parameters and cognitive outcomes. The study specifically investigates not only the individual effects of hemodynamic parameters such as mean arterial pressure, peripheral oxygen saturation (SpO₂), invasive blood pressure values (systolic and diastolic), and heart rate but also explores the predictive value of their mathematically derived combinations for neurocognitive dysfunction. Specific objectives include: (1) developing predictive models for postoperative neurocognitive dysfunction using detailed intraoperative hemodynamic data; (2) identifying optimal ranges for hemodynamic parameters that minimize neurocognitive risk; (3) evaluating the relative importance of different hemodynamic parameters and their combinations in predicting cognitive outcomes; (4) comparing the performance of traditional statistical methods with advanced machine learning algorithms; and (5) providing evidence-based recommendations for intraoperative hemodynamic management to reduce postoperative cognitive complications. This prospective observational study included 28 patients undergoing elective CABG surgery. A total of 730,870 hemodynamic data points were collected during surgery, obtained from the anesthesia machine at regular intervals. The hemodynamic parameters monitored included mean arterial pressure (MAP), peripheral oxygen saturation (SpO₂), invasive blood pressure values (systolic and diastolic), and heart rate (PLS). Additionally, patient demographic information (age, body surface area), pump flow rate, bypass duration, and cross-clamp time were recorded and included in the analysis. Neurocognitive function was evaluated using the Montreal Cognitive Assessment (MoCA) test, a validated and widely used screening tool for detecting cognitive impairment. The test was administered preoperatively by anesthesiology and reanimation specialists, and postoperatively (at one month) by cardiovascular surgery specialists. A decrease of 2 or more points in the MoCA score was defined as neurocognitive dysfunction, following established criteria in the literature. In this cohort, postoperative neurocognitive dysfunction was observed in 36% of patients (n=10). Beyond analyzing raw hemodynamic parameters, several mathematically derived features were calculated based on established threshold values. These included: (1) Absolute Maximum Drop (AMD) - the maximum deviation below threshold values of 75, 65, and 55 mmHg; (2) Time Under Threshold - the duration spent below each threshold value; and (3) Area Under Threshold - the integral of the area below each threshold value over time. These derived parameters were calculated for different surgical phases: entire operation duration, bypass period, and cross-clamp period. Multiple analytical approaches were employed to develop predictive models for neurocognitive dysfunction. Logistic regression was used as a baseline statistical method, providing interpretable coefficients and odds ratios for individual parameters. Multilayer Perceptron (MLP) neural networks were implemented to capture non-linear relationships between hemodynamic parameters and cognitive outcomes. The MLP architecture utilized ReLU activation functions in hidden layers, sigmoid activation in the output layer, Xavier weight initialization, and cross-entropy loss function optimized through backpropagation. Random Forest algorithm emerged as the most effective approach for this classification problem. Random Forest, an ensemble learning method, creates multiple decision trees during training and outputs the mode of their predictions. The algorithm's ability to handle complex interactions between variables, provide feature importance rankings, and generate partial dependence plots (PDPs) made it particularly valuable for identifying optimal hemodynamic parameter ranges. Cross-validation was employed to prevent overfitting and ensure model generalizability. Python programming language was used for implementing machine learning algorithms, while IBM SPSS Statistics Version 26 was utilized for statistical analyses. The Random Forest algorithm demonstrated superior performance compared to other methods, achieving accuracy rates up to 90% in predicting postoperative neurocognitive dysfunction. Specifically, models utilizing combinations of heart rate (PLS), peripheral oxygen saturation (SpO₂), systolic blood pressure (IBPS), and diastolic blood pressure (IBPD) yielded the most successful predictions. The ROC-AUC values ranged from 0.83 to 0.90 across different surgical phases and parameter combinations, indicating excellent discriminative ability. Cross-validation results confirmed the robustness of these models, with consistent performance across different patient subsets.