Prediction of Postoperative Delirium in Older Adults from Preoperative Cognition and Occipital Alpha Power from Resting-State Electroencephalogram

Abstract

Background: Postoperative delirium is the most common complication following surgery among older adults, and has been consistently associated with increased mortality and morbidity, cognitive decline, and loss of independence, as well as markedly increased health-care costs. Electroencephalography (EEG) spectral slowing has frequently been observed during episodes of delirium, whereas intraoperative frontal alpha power is associated with postoperative delirium. We sought to identify preoperative predictors that could identify individuals at high risk for postoperative delirium, which could guide clinical decision-making and enable targeted interventions to potentially decrease delirium incidence and postoperative delirium-related complications.

Methods: In this prospective observational study, we used machine learning to evaluate whether baseline (preoperative) cognitive function and resting-state EEG could be used to identify patients at risk for postoperative delirium. Preoperative resting-state EEGs and the Montreal Cognitive Assessment were collected from 85 patients (age = 73 ± 6.4 years, 12 cases of delirium) undergoing elective surgery. The model with the highest f1-score was subsequently validated in an independent, prospective cohort of 51 older adults (age = 68 ± 5.2 years, 6 cases of delirium) undergoing elective surgery.

Results: Occipital alpha powers have higher f1-score than frontal alpha powers and EEG spectral slowing in the training cohort. Occipital alpha powers were able to predict postoperative delirium with AUC, specificity and accuracy all >90%, and sensitivity >80%, in the validation cohort. Notably, models incorporating transformed alpha powers and cognitive scores outperformed models incorporating occipital alpha powers alone or cognitive scores alone.

Conclusions: While requiring prospective validation in larger cohorts, these results suggest that strong prediction of postoperative delirium may be feasible in clinical settings using simple and widely available clinical tools. Additionally, our results suggested that the thalamocortical circuit exhibits different EEG patterns under different stressors, with occipital alpha powers potentially reflecting baseline vulnerabilities.

Figure 1. Schematic diagram of machine learning framework.

A) Data exploration. A1) different band powers of different regions of interested (ROIs) were extracted from SAGES EEG Data Set. A2) the results are visualized using scatter plots as well as EEG topographic plots. A3) occipital alpha powers (and sub-alpha powers) from both eyes-open and eyes-closed condition was selected. Baseline cognition (MoCA) was selected a priori. B) Model Selection. The feature sets consisting of EEG alpha (and sub-alpha) powers determined from data exploration (A), along with MoCA selected a priori, were tested using 9 different classifiers using cross-validation (red arrows). The details of the cross-validation is shown in (D). C) After the classifier with the highest f1-score is determined in model selection step, the classifier was re-trained on the entire SAGES Data Set, then the parameters of the classifier were held fixed and independently validated on INTUIT/PRIME Data Set. The results are plotted in Figure 4 and tabulated in Table 3. D) Details of the cross-validation used in model selection step. The CV performance are plotted in Supplementary Figure S6 and Supplementary Table S6.

Figure 2. Scatter plots of individual EEG alpha powers in O2 channel.

Scatter plot showing distributions of alpha powers in participants with post-operative delirium (purple dots) and without delirium (blue dots). Black squares represent the mean of their respective groups. Left panel shows the untransformed alpha powers as seen in PSD plots (Fig. 3) and right panel shows transformed alpha powers. Values enclosed in parentheses in the x axis represent the p-values of two-sample two-tailed Welch’s t-test. This channel is representative of all channels in the occipital region and their scatter plots can be found in the Supplementary Fig. S2 & S3.

Figure 3. Power spectral densities of O2 channel.

Power spectral densities of O2 channel from both SAGES and INTUIT/PRIME Data Sets in both eyes-open (EO) and eyes-closed (EC) conditions for [1, 26] Hz frequency range. Shaded regions represent standard error of the means. Blue represents control group and orange represents delirium group. PSDs for O1 and POz are shown in Supplementary Figure S4 & S5, respectively.

Figure 4. Model validation: performances using the Principal Feature Set (Alpha Powers + MoCA).

A) The performance of the model (LDA LW) selected from the model selection step. 95% confidence intervals are shown as thin vertical bars. Chance levels for each metric are shown as dark horizontal lines (details of chance levels in Supplementary Materials). Blue: accuracy, orange: sensitivity, green: specificity, red: f1 score, purple: AUC of ROC curve, brown: AUC of precision-recall curve, pink: positive predictive value (PPV, also known as precision), grey: negative predictive value (NPV). Values are tabulated in Table 4A. ROC (B) and PR (C) curves are shown for the selected LDA LW model. The AUCs reported