An evolutionary spiking neural network developed to predict postoperative delirium achieved modest diagnostic performance and did not outperform standard histogram gradient boosting, according to a research study published in Frontiers in Neurology. The investigation, featured in the journal’s Artificial Intelligence in Neurology section, evaluated whether combining an evolutionary algorithm with surrogate-gradient training could improve risk prediction from preoperative hospital records. When tested on surgical patients to predict physician-diagnosed delirium within their first postoperative week, the specialized neural network architecture underperformed an established histogram gradient boosting baseline on primary discrimination benchmarks.
Postoperative delirium is recognized as a serious clinical complication that can occur after surgery. To investigate machine learning approaches for perioperative risk identification, a research team led by Haohan Zhang, Yong Yang, Lei Zhao, Jianyu Hao, Xutao Bian, and Ke Cui evaluated a shared evolutionary plus surrogate-gradient spiking neural network. The collaborating investigators, representing Xuanwu Hospital Capital Medical University, the Chinese Academy of Sciences Chongqing Institute of Green and Intelligent Technology, Chongqing University of Posts and Telecommunications, and Rongcheng County People’s Hospital, examined preoperative hospital records alongside a separate exploratory intensive care unit dataset.
The primary clinical investigation evaluated a hospital cohort consisting of 577 orthopedic surgery patients. Within this patient group, 84 individuals had physician-documented delirium recorded during the first postoperative week following their orthopedic procedures. To assess model stability and evaluate predictive performance without bias, the researchers divided the clinical records into three stratified splits. This evaluation framework allowed the team to systematically test the evolutionary spiking neural network across multiple partitioned subsets of the orthopedic surgery cohort.
To prepare the clinical data for predictive modeling, the researchers implemented a structured training pipeline that relied entirely on training-only preprocessing. Feature selection was conducted using LASSO regression methods to identify relevant predictive variables from preoperative patient records. Additionally, minority oversampling was applied to mitigate the class imbalance between patients with physician-documented delirium and those without the condition. Both the hospital orthopedic cohort and the comparative baseline models followed this exact preprocessing, feature selection, and oversampling pipeline.
Across the three stratified hospital splits, the evolutionary spiking neural network demonstrated modest discriminative ability for predicting delirium within the first postoperative week. The model achieved a mean area under the receiver operating characteristic curve of 0.629, with a standard deviation of 0.075. The network also recorded a mean average precision of 0.278, with a standard deviation of 0.140. When assessed at the common decision threshold of 0.50, the model attained a mean positive-class F1 score of 0.279, with a standard deviation of 0.167. Across both the area under the receiver operating characteristic curve and average precision metrics, histogram gradient boosting demonstrated superior performance compared to the spiking neural network.
In parallel with the orthopedic surgical cohort, the authors analyzed the MIMIC-III database as a distinct exploratory intensive care unit classification task. Both datasets utilized training-only preprocessing, LASSO feature selection, minority oversampling, and the identical baseline pipeline. However, the investigators explicitly noted that the MIMIC-III findings were exploratory in nature and do not constitute external validation for postoperative delirium prediction. The authors emphasized that the intensive care unit analysis represents a separate task rather than confirmatory evidence of the model’s perioperative clinical utility.
In their final assessment, the researchers concluded that the evolutionary spiking neural network demonstrated modest, dataset-specific performance and did not outperform histogram gradient boosting on the main discrimination measures. The authors reiterated that the MIMIC-III analysis should be interpreted strictly as exploratory intensive care unit classification, rather than external validation for postoperative delirium prediction after orthopedic surgery. The peer-reviewed study was accepted for publication in November 2026 after initial submission in August 2026.
Subject of Research: Predicting postoperative delirium using an evolutionary spiking neural network
Article Title: Research on perioperative delirium prediction using evolutionary spiking neural networks
Article References: Zhang, H., Hao, J., Bian, X., Cui, K., Yang, Y., & Zhao, L. (2026). Research on perioperative delirium prediction using evolutionary spiking neural networks. Frontiers in Neurology, 17, 1970193. https://doi.org/10.3389/fneur.2026.1970193
Image Credits: AI Generated
DOI: 10.3389/fneur.2026.1970193
Keywords: postoperative delirium, spiking neural networks, evolutionary algorithms, orthopedic surgery, machine learning, histogram gradient boosting, predictive modeling, preoperative hospital records, LASSO feature selection, minority oversampling, AUROC, average precision, MIMIC-III, intensive care unit classification, Frontiers in Neurology, surrogate gradient, perioperative monitoring, clinical risk prediction
News Source: Cassandra Pierce. (November 20, 2026). Spiking Neural Network Shows Modest Accuracy in Predicting Postoperative Delirium. Scienmag.



