Descriptive statistics for: Applying a transformer architecture to intraoperative temporal dynamics improves the prediction of postoperative delirium

Background. Patients who experienced postoperative delirium (POD) are at higher risk of poor outcomes like dementia or death. Previous machine learning models predicting POD mostly relied on time-aggregated features. We aimed to assess the potential of temporal patterns in clinical parameters during surgeries to predict POD.

Methods. Long short-term memory (LSTM) and transformer models, directly consuming time series, were compared to multi-layer perceptrons (MLPs) trained on time-aggregated features. We also fitted hybrid models, fusing either LSTM or transformer models with MLPs. Univariate Spearman’s rank correlations and linear mixed-effect models establish the importance of individual features that we compared to transformers’ attention weights.

Results. We found that best performance is achieved by a transformer architecture ingesting 30 minutes of intraoperative parameter sequences. Systolic invasive blood pressure and given opioids mark the most important input variables, in line with univariate feature importances.

Conclusion. Intraoperative temporal dynamics of clinical parameters, exploited by a transformer architecture named TRAPOD, are critical for the accurate prediction of POD

Methods
We identified promising features due to a literature review and found a potential number of 197 features in the clinical information systems (CIS) across three different hospital sites of our center (see Table B.3 in Supplement B). We selected 148 out of 197 variables due to their availability for at least 1% of patients. Thus, we investigated the influence of rare as well as highly (100%) available features. Details on feature availability (and missingness) are provided in Table B.4 in Supplement B. Table 3 summarizes the feature encoding process. Feature values were either considered as time-static, not changing over the intraoperative phase, or time-dynamic, fluctuating during the surgery. In addition to 148 selected features, we derived four composite features that combined 1. non-invasive and invasive mean blood pressure, 2. set and measured fraction of inspired oxygen (FiO2), 3. invasive and spontaneous urine output, 4. set and measured positive end-expiratory pressure (PEEP). Single feature vectors were simply concatenated for these pooled measures before sampling with an interval of e.g. three minutes. We introduced four composite features to increase data availability for these variables, as they depict the same physiological attributes such as blood pressure. By keeping the original single vectors in our feature set, we could differentiate e.g. between spontaneous and mechanical ventilation. For 19 medications, the cumulative sum of administered volumes or amounts over time was calculated. In addition to these derived variables, we encoded data availability with binary missingness indicators for 67 features, assigning 1 if a value was missing and 0 otherwise79. Binary missingness indicators were included for the following clinical domains: EEG (5 features), inputs (19 features), outputs (3 features), laboratory values (8 features), scores (4 features), vital signs (12 features), respiratory signals (8 features), demographics (5 features excluding gender), and four composite features. For other domains, like medical history, we could not differentiate between a missing measurement (variable not present) and a true negative (variable encodes a negative result). Thus, no binary missingness indicators were added here. A total of 238 features were included in our final feature set (see Table B.5 in Supplement B)