Transformer representation learning is necessary for dynamic multi-modal physiological data on small-cohort patients

TL;DR

This study tackles early Detection of postoperative delirium (POD) in a small, multi-modal ICU cohort by introducing a Transformer-based representation learning framework tailored to dynamic physiological signals. The proposed Fusion Pathformer, an adaptation of Pathformer for multi-modal data with a TrendLoss regularizer, learns unified temporal representations that markedly improve POD prediction over traditional classifiers, achieving AUROC values above $0.95$ for POD$\_2$ and POD$\_3$ in TYPE I patients. TYPE II results indicate temporal-dimension transformers can enhance sensitivity and specificity when inter-modal relations are less informative, while patch-based multimodal designs show trade-offs between sensitivity and specificity. The work highlights the critical role of representation learning for multi-modal medical time series in POD diagnosis, while acknowledging limitations from small sample size and data heterogeneity, and calls for public, larger-scale datasets to enable robust clinical deployment.

Abstract

Postoperative delirium (POD), a severe neuropsychiatric complication affecting nearly 50% of high-risk surgical patients, is defined as an acute disorder of attention and cognition, It remains significantly underdiagnosed in the intensive care units (ICUs) due to subjective monitoring methods. Early and accurate diagnosis of POD is critical and achievable. Here, we propose a POD prediction framework comprising a Transformer representation model followed by traditional machine learning algorithms. Our approaches utilizes multi-modal physiological data, including amplitude-integrated electroencephalography (aEEG), vital signs, electrocardiographic monitor data as well as hemodynamic parameters. We curated the first multi-modal POD dataset encompassing two patient types and evaluated the various Transformer architectures for representation learning. Empirical results indicate a consistent improvements of sensitivity and Youden index in patient TYPE I using Transformer representations, particularly our fusion adaptation of Pathformer. By enabling effective delirium diagnosis from postoperative day 1 to 3, our extensive experimental findings emphasize the potential of multi-modal physiological data and highlight the necessity of representation learning via multi-modal Transformer architecture in clinical diagnosis.

Figures (6)

• Figure 1: Overview of study design. Post-operative delirium, a prevalent acute neuropsychiatric syndrome is crucial and challenging for early diagnosis. A cascading structure of predictive system was established with components of data preprocessing, Transformer representation learning and machine learning classifier, utilizing multi-modal physiological data including aEEG, vital signs, ECG and hemodynamics. Based on Pathformerchen2024pathformermultiscaletransformersadaptive, we proposed fusion adaptation version for multi-modal physiological data and introduced a new regularization item TrendLoss to assist with representation learning. We investigated the average prediction performance of post-operative day 1 to 3 with several Transformer representations. AUROC, sensitivity and representation visualization exhibited a profound improvement with our Fusion Pathformer.
• Figure 2: Visualization of training set using t-SNE dimension reduction. Positive samples came from patients diagnosed POD on the first post-operative day while negative samples came from those diagnosed normal. We generated the samples by window sliding, the total sample size is 1824 with prediction period of 30 minutes.
• Figure 3: Impact of TrendLoss regularization strength $\lambda$. Different results on POD1-3 emphasizes different optimal choices for different POD indicators and different machine learning classifiers.
• Figure 4: Impact of varying prediction period $T$. Results of Logistic Regression and SVM under all POD indicators exhibit the similar pattern when prediction period $T$ is changing. Subfigure (A)(B)(C) indicate that the diagnosis of POD patients is quite sensitive to the prediction period $T$. These results emphasize that macrospoic information and longer monitoring duration are critical for POD diagnosis.
• Figure 5: Our design of representation-based post-operative delirium prediction model. The model comprises three main parts, data preprocessing, representation learning and predictive classification learning.