This study aims to predict in-hospital 30-day postoperative mortality using preoperative anesthesia assessments. Data were collected from the electronic health system of the Far Eastern Memorial Hospital, a large academic medical center in Taiwan. Preoperative anesthesia assessment records and discharge summaries were included. Overall, 5 years’ worth of retrospective data were collected from January 1, 2016, to December 30, 2020. The last version of the anesthesia assessment was included for each surgery. Patients over 18 years of age who underwent at least 1 surgical procedure under general or neuraxial anesthesia were included. Cases with an ASAPS of 6 were excluded. Records lacking entry time, exit time, preoperative diagnosis, or proposed procedure text were excluded. The in-hospital 30-day postoperative mortality was defined by a discharged route of “expired” and “critical against-advice discharge” (when the patient wants to die at home) without future admission. Discharges within 30 days after surgery were identified and labeled as “true”; those occurring outside this window were marked as “false.” The end date of the testing set was November 30, 2020, 30 days before the end of the collected data, to ensure complete 30-day mortality detection (Figure 1).

The Institutional Review Board of the Far Eastern Memorial Hospital approved this retrospective study and waived the requirement of informed consent (#109129-F and #110028-F).

We collected 123,718 surgery results for patients aged over 18 years. After applying the exclusion criteria, a cohort of 123,515 (99.8%) patients who underwent surgeries remained. A final cohort of 121,313 (98.2%) patients was used after removing those who underwent surgeries after November 30, 2020 (Figure 1). The training, validation, and testing cohorts finally contained 79,324 (68.7%), 19,832 (17.2%), and 16,267 (14.1%) of 115,423 patients. Patient characteristics of the training, validation, and testing cohorts are listed in Table 1. In the overall cohort, most patients had an ASAPS of 2 or 3. Overall, 107,176 (88.5%) of patients were under general anesthesia. The most prevalent comorbidities were hypertension (n=43,391, 35.8%), followed by diabetes (n=24,314, 20.0%). A total of 1562 (1.3%), 997 (1.3%), 249 (1.3%), and 215 (1.3%) patients died within 30 days of surgery in the overall, training, validation, and testing cohorts, respectively. Multimedia Appendix 1 present a summary of the laboratory data and preoperative vital signs.

The input features included patient characteristics (age, height, weight, BMI, sex, ASAPS, ASA emergency status, department, preoperative location, and anesthesia type), surgery characteristics (emergency level, preoperative diagnosis, and proposed procedure), comorbidities (diabetes mellitus, hyperlipidemia, hypertension, cerebrovascular accident, cardiac disease, chronic obstructive pulmonary disease, asthma, hepatic disease, renal disease, bleeding disorder, major operations, smoking, and drug allergy), preoperative laboratory data (hemoglobin, platelet, international normalized ratio, prothrombin time, activated partial thromboplastin time, creatinine, aspartate transaminase, alanine transaminase, blood sugar, serum sodium, and serum potassium), and preoperative vital signs (body temperature, oxygen saturation, heart rate, respiratory rate, systolic and diastolic blood pressure, and consciousness status); see Table 2.

Continuous features (eg, age, height, weight, latest laboratory data before surgery, and preoperative vital signs) were standardized by subtracting the mean and scaling to variance. Outliers were regarded as input errors and treated as missing data. Multimedia Appendix 2 lists the definitions of the outliers. Missing values were imputed with the median value of the data set for continuous features.

Categorical features with only 2 classes (eg, sex, comorbidities, ASA emergency status, and consciousness status) were converted into binary encoding. All other categorical features (eg, ASAPS [5 classes], department [22 classes], emergency level [4 classes], preoperative location [4 classes], and anesthesia type [4 classes]) were transformed into one-hot encodings. Missing data were imputed with the majority category of the training data set. The preoperative diagnoses and proposed procedures were expressed as free text. Characters other than alphabetical and numerical ones were removed (eg, Chinese characters [typically notes for colleagues only] and punctuation). English stop words providing no helpful information to the model (eg, “a,” “in,” and “the”) were removed using the Natural Language Toolkit [20].

We used the previous 4 years’ surgery results to predict the last year results. Patients who underwent surgeries between January 1, 2016, and December 31, 2019, were selected and split into training and validation sets in a 4:1 ratio; those who underwent surgeries between January 1, 2020, and November 30, 2020, were selected as the testing set (Figure 1). Patients in the training or validation set were removed from the testing set to prevent information leakage [6].

Our results were compared with state-of-the-art models, using patient preoperative vital signs and laboratory data to predict in-hospital 30-day mortality [6]. Meanwhile, to demonstrate the effect of adding preoperative diagnoses and proposed procedures to the prediction model, we added text features and compared the performances of the highest-performing models.

First, we compared the state-of-the-art models using patient and surgery characteristics (without text), comorbidities, preoperative vital signs, and laboratory data to predict the in-hospital 30-day mortality. Figure 2B shows our proposed DNN model with 4 fully connected (FC) layers and a Softmax layer output function. We compared our DNN model with other ML models, including a random forest classifier (with 2000 estimators and Gini impurity as the splitting criterion) [21], extreme gradient boosting (XGBoost, with a learning rate of 0.3 and a maximum depth of 6) [22], and logistic regression (with an L2 penalty); see Figure 2A. To balance the data while training the ML models, oversampling by 78 times was performed on the training set via the synthetic minority oversampling technique; this produced synthetic samples along a straight line between randomly selected samples in the feature space [23]. While training our DNN model, we adjusted the weight to compensate for the imbalanced classes. We added the text of preoperative diagnoses and proposed procedures to the DNN model architecture (denoted as BERT-DNN; see Figure 2C) and compared its performance with those of other models.

The language model extracted features from the preprocessed text. Figure 2C shows the architecture of the language model. The preprocessed texts were tokenized using the BERT tokenizer, which transformed each word fragment into a unique token designed for use in BERT’s pretraining process [17]. Then, these tokens were embedded by Bio+Clinical BERT, a variant of BERT pretrained on text from PubMed and Medical Information Mart for Intensive Care III [24]. The text information was transformed into a 768-dimension vector (the “word embeddings”) at the pooler output layer [17,24]. These word embeddings were input into 2 FC layers before concatenation with other structured features. The concatenated vectors were input into 3 FC layers and a Softmax layer output function. Figure 2C shows the architecture of the BERT-DNN model.

Cross-entropy was used as the loss function. Class weight imbalances were compensated for by setting the weights as the inverses of the different classes’ frequencies (~1:78). Further, the training data were split into training and validating sets in a 4:1 ratio to train the deep learning (DL) model. We used AdamW from the PyTorch package as the optimizer, setting a learning rate of 0.00002 for both DL models. We trained our BERT-DNN and DNN models with batch sizes of 64 and 512, respectively, until the 100th epoch. The DL model with the smallest validation loss was selected for performance comparison.

The models were evaluated using the AUROC, the area under the precision-recall curve (AUPRC), sensitivity (also referred to as recall), specificity, precision (also called the positive predictive value), and the F1 score. The F1 score was a harmonic mean of recall and precision and was calculated as 2/[(1/recall) + (1/precision)]. Because postoperative mortalities accounted for 1.3% (1562/121,313) of our data set, classes were extremely imbalanced between the positive and negative groups. Here, the AUPRC (which calculated the average precision) was better than the AUROC for evaluating the discrimination of models [25,26]. For comparison of AUROCs, we applied a nonparametric approach proposed by DeLong et al [27] to calculate the SE of the area and the P value. P<.05 was regarded statistically significant. We calculated exact binomial 95% CIs for the AUROC. For comparison of AUPRCs, we performed bootstrapping 1000 times in the testing set to calculate the difference in areas and the 95% CI [28]. If the 95% CI for the difference in areas does not include 0, it can be concluded that these 2 areas are significantly different (P<.05). We performed bootstrapping 1000 times in the testing set to calculate the 95% CI for other metrics [6]. The predicted probabilities were calibrated using the histogram bins technique, using the same observed mortality in each bin of the validation set [8]. After calibration, the mean observed incidences of mortality were plotted against the mean predicted probabilities within groups in the testing set.

To show the correlation between increased prediction probabilities and text inputs, the t distributed stochastic neighbor embedding (SNE) was implemented by reducing the 768 dimensions of the language model’s pool output to 2 into a plane [29,30]. Thus, we showed the clustering of word embeddings using assorted colors for different predicted probabilities and different icons for observed mortalities. We randomly resampled 10,000 and 5000 patients who underwent surgeries in the training and testing sets, respectively, to construct this visualization. The language-model-predicted probabilities and observed mortalities for randomly selected text inputs were calculated and listed.

The study was implemented using Python 3.9, Scikit-learn 0.24 [31], imbalanced-learn 0.8.0 [23], PyTorch 1.8 [32], and transformers 4.9 (Hugging Face) [24]. Our models were trained and validated on the NVIDIA Tesla P100-PCIE-16GB graphics processing unit (GPU). The statistical significances of AUROCs and AUPRCs were calculated using MedCalc software (Ostend, Belgium).

The BERT-DNN had the highest AUROC of 0.964 (95% CI 0.961-0.967) and the highest AUPRC of 0.336 (95% CI 0.276-0.402); see Table 3 and Figure 3. The random forest achieved the second-highest AUROC of 0.961 (95% CI 0.958-0.964), and the DNN achieved the second-highest AUPRC of 0.319 (95% CI 0.260-0.384). The BERT-DNN model had the highest F1 score of 0.347 (95% CI 0.305-0.388).

The BERT-DNN had a significantly higher AUROC compared to XGBoost, logistic regression, and ASAPS but not a significantly higher AUROC compared to the DNN and the random forest (Table 4). The BERT-DNN also had a significantly higher AUPRC compared to the DNN, random forest, XGBoost, logistic regression, and ASAPS (Table 5).

In the BERT-DNN model, when the predicted probability of mortality increased from 0.2% to 39.4%, the observed incidence increased from 0.2% to 42.7% (Figure 4).

Because the observed mortalities were distributed concordantly with increased prediction probabilities, the annotated scatter plots showed that the text contributed to low- and high-probability predictions (Figure 5, Multimedia Appendix 3). Table 6 lists the probabilities predicted by the language model and the mortalities observed for a randomly selected text input.

The DNN-BERT model predicted the in-hospital 30-day mortality with the highest AUROC of 0.964 (95% CI 0.961-0.967) and an AUPRC of 0.336 (95% CI 0.276-0.402); see Table 3 and Figure 3. The BERT-DNN had an AUROC significantly higher compared to XGBoost, logistic regression, and ASAPS but not the DNN or random forest. The BERT-DNN also had an AUPRC significantly higher compared to the DNN, random forest, XGBoost, logistic regression, and ASAPS.

Hill et al [6] proposed an ML model that outperformed previous tools (eg, preoperative score to predict postoperative mortality, Charlson comorbidity, and ASAPS) and could be used independently by clinicians. Our BERT-DNN model outperformed Hill et al’s [6] model, obtaining a higher AUROC, sensitivity, and F1 score than their results (0.964, 95% CI 0.961-0.967 vs 0.932, 95% CI 0.910-0.951; 0.650, 95% CI 0.587-0.719 vs 0.239, 95% CI 0.127-0.379; and 0.347, 95% CI 0.305-0.388 vs 0.302, 95% CI 0.172-0.449, respectively); see Table 3. The preoperative diagnosis text features and proposed procedure information might contribute to our BERT-DNN model and enhance its sensitivity and F1 score. Unlike Hill et al [6], who focused on patients undergoing general anesthesia, we trained and tested our model on both general and neuraxial anesthesia. The DL model with clinical text predicted postoperative mortality significantly more discriminatively than logistic regression and ASAPS (Table 4).

DL methods predict postoperative mortality using preoperative and intraoperative features [7-9]. Using a summary of intraoperative features alongside the ASAPS, Lee et al [7] presented a DNN model that achieved an AUROC of 0.91 (95% CI 0.88-0.93). Our DNN model obtained a higher AUROC than their model because we included key features such as preoperative location and surgical department, the importance of which was also verified in previous studies [6]. Fritz et al [8] proposed a multipath convolutional neural network model to predict postoperative mortality using intraoperative time-series data and preoperative features. Their model achieved an AUROC of 0.910 (95% CI 0.897-0.924) and an AUPRC of 0.325 (95% CI 0.280-0.372) [33]. In contrast, our model can be used preoperatively and achieve a higher AUROC and AUPRC (Table 3).

Previous studies used ICD and CPT codes as categorical features to stratify surgery risk [2,6,7,9,12]. This input feature has many classes, which resulted in a sparse input matrix; this made it difficult for the model to learn helpful information. However, because ICD codes were typically recorded after surgery, including them in the preoperative model was impractical. Furthermore, the CPT code was not used globally. For this reason, we could not compare a model including word embeddings with one including CPT codes. However, our results exhibited excellent discrimination with a high AUROC and AUPRC. The AUPRC is significantly higher than models without text. The calibration plot also strongly correlated the predicted probabilities and observed mortalities (Figure 4). Word embedding visualizations showed that the increased predicted probabilities were concordant with high-risk surgery and an increased mortality rate (Figure 5 and Table 6). We showed that word embeddings for surgery information could be used in DL models to predict postoperative mortality before surgery without requiring CPT or ICD codes.

The fusion of neural networks, combining diverse types of data (eg, image [34] and time-series [8] data) with 1D data (eg, categorical, and continuous data), improved the model’s performance. Including unstructured clinical text via natural language procession can improve intensive care unit (ICU) mortality predictions [14,16]. The DL model that combined unstructured and structured data outperformed models using either type of data alone [15]. Moreover, the performance of the clinical pretrained DL language model could be maintained between different institutions [35].

Our study has several limitations. First, postoperative mortality accounted for 1.3% (1562/121,313) of our cohort, and the classes were highly imbalanced. The model training and performance metric evaluations were difficult to apply with these sparse positive labels. To compensate for the class imbalance via an algorithmic method, we applied cost-sensitive learning by balancing the weights of the loss function to emphasize the minority group [36]. We evaluated the discrimination of our model with the AUPRC, which is more informative than the AUROC for imbalanced data [8,25,26]. Second, our model predicted mortality using EHRs. The errors in the records and missing values affected the prediction results. Typos of text interfered with the word-embedding process. Outliers were detected and input using the defined rules (Multimedia Appendix 2). Third, all records were collected from a single large medical center. Although the pipeline we created ensured that the DL model could be reproduced in other institutes, the model weights might vary for a different data set. The generalizability of our results must be examined in future studies.

In conclusion, descriptive surgical text was essential for predicting postoperative mortality. The word embeddings of preoperative diagnoses and proposed procedures, via the contextualized language model BERT, were combined in DL models to predict postoperative mortality. This predictive capacity can help identify patients with higher risk from structure data and text of EHRs.