Predicting the causes of unplanned reoperations using an AI-based multimodal system: a multicentre, multidepartmental study based on EMRs Authors: Luyuan Xie, Congpu Zhao, Xutong Tan, Pengyu Zhao, Xuan Gong, Shengyang Li, Qingni Shen, Tianyu Luan, Guochen Yan, Dan Wang, Jiong Zhou, Xiaojun Ma, Zhonghai Wu Published: 2026 ABSTRACT ObjectiveUnplanned reoperations are widely recognised as key indicators of surgical quality and patient safety. Analysing the underlying causes of such reoperations is critical for guiding clinical decision-making and optimising patient management. Most existing studies relied on electronic medical records (EMRs) for simple retrospective analyses and focused exclusively on single-disease cohorts, limiting their generalisability. Tools capable of predicting the causes of unplanned reoperations using multimodal EMRs across diverse clinical settings remain scarce. This study aimed to develop an artificial intelligence (AI)-based multimodal system integrating structured and unstructured EMR data across institutions.Methods and analysisWe developed the Multi-modal Prediction System for Causes of Unplanned Reoperation (MPSUR), an AI-based framework trained on a retrospective cohort of 2922 cases collected from 15 departments across 8 hospitals (2015–2024). The system integrates structured series data (such as age and sex) and clinical text (such as diagnoses and procedures) using the Graph Convolutional Network and Time and Frequency Recurrent Neural Network. Modality-specific and fused predictions were generated using linear classifiers.ResultsIn the clinical reader study, the MPSUR outperformed surgeons across all departments (mean accuracy=60.27%, 95% CI 58.27% to 62.26%). Feature ablation showed that variables such as procedure and department were the most predictive, with their removal significantly reducing performance (−3.05%, p<0.01) for the internal dataset and (−3.42%, p<0.005) for the external dataset. MPSUR achieved 62.03% and 56.41% accuracy on internal and external datasets, respectively, outperforming classical baselines by up to 8%. The SD across cross-validation folds was minimal, underscoring model stability. It also remained robust when only text or series data were available.ConclusionThe MPSUR provides an automatic, accurate, interpretable and generalisable tool for predicting the causes of unplanned reoperations using multimodal EMRs. Its strong performance across internal and external datasets, as well as its superiority over clinicians, supports its clinical utility as a decision-support system to enhance surgical safety and patient outcomes. WHAT IS ALREADY KNOWN ON THIS TOPIC Most studies of unplanned reoperation are retrospective and confined to single-disease cohorts, with no available tool to predict causes using multimodal electronic medical records (EMRs) across diverse clinical settings, thereby limiting generalisability. Most studies of unplanned reoperation are retrospective and confined to single-disease cohorts, with no available tool to predict causes using multimodal electronic medical records (EMRs) across diverse clinical settings, thereby limiting generalisability. WHAT THIS STUDY ADDS This study introduced Multi-modal Prediction System for Causes of Unplanned Reoperation (MPSUR), an artificial intelligence-based framework that integrates structured series and unstructured text from EMRs to predict the causes of unplanned reoperation. Trained on a multicentre and multidepartmental dataset and externally validated, the MPSUR significantly outperformed experienced surgeons across departments and traditional machine learning models. This study introduced Multi-modal Prediction System for Causes of Unplanned Reoperation (MPSUR), an artificial intelligence-based framework that integrates structured series and unstructured text from EMRs to predict the causes of unplanned reoperation. Trained on a multicentre and multidepartmental dataset and externally validated, the MPSUR significantly outperformed experienced surgeons across departments and traditional machine learning models. HOW THIS STUDY MIGHT AFFECT RESEARCH, PRACTICE OR POLICY The MPSUR provides an automated, robust and generalisable decision-support tool for identifying the causes of unplanned reoperations, maintaining effectiveness even with incomplete data modalities, thereby enhancing its applicability in real-world settings. This lays the groundwork for intelligent surgical quality monitoring in hospitals, particularly where manual reviews are resource intensive. The MPSUR provides an automated, robust and generalisable decision-support tool for identifying the causes of unplanned reoperations, maintaining effectiveness even with incomplete data modalities, thereby enhancing its applicability in real-world settings. This lays the groundwork for intelligent surgical quality monitoring in hospitals, particularly where manual reviews are resource intensive. Introduction Unplanned reoperation,1 defined as the need for an additional unplanned procedure during hospital admission, is a well-established indicator of surgical quality and patient safety.2 In clinical practice, enhanced recognition of perioperative complications, particularly severe complications requiring reoperation in patients who have undergone their first surgery, is essential for improving outcomes. However, previous studies have shown that baseline patient conditions, medication use and intraoperative events are closely associated with unplanned reoperations. Consequently, the underlying causes can often be identified retrospectively, and reliable approaches for early clinical warning are lacking. Accordingly, the development of a system capable of predicting the causes of unplanned reoperation before surgery is critically important for informing subsequent clinical decision-making and enhancing patient management. Despite growing interest in unplanned reoperations, current research predominantly focuses on statistical analysis of electronic medical records (EMRs)3 4 and tends to concentrate on single disease types.5 6 These studies leveraged EMRs7 8 which contain rich and structured clinical data related to surgery together with traditional statistical methods to assess incidence or identify associated risk factors. However, two limitations remain in predicting the causes of unplanned reoperations. First, although substantial EMR data have been collected, studies have been limited to retrospective analyses and widely applicable predictive systems with good interpretability and generalisability across diverse clinical scenarios are lacking, restricting their use in real-world settings for unplanned reoperations. Second, prior research has focused mainly on single diseases, primarily because such events have low incidence and their records are dispersed across different diseases and surgical pathways, resulting in a lack of large-scale structured datasets covering multiple diseases. To address the aforementioned challenges, we constructed a comprehensive dataset of unplanned reoperations based on EMRs collected from 15 departments across 8 hospitals, covering the period from 2015 to 2024. Leveraging this multicentre and multidepartmental EMR dataset, we developed the Multi-modal Prediction System for Causes of Unplanned Reoperation (MPSUR), an artificial intelligence (AI)-based framework that integrates structured series data9 and clinical text data10 11 for accurate and generalisable prediction. Comparative experiments conducted across both unimodal and multimodal datasets from internal and external institutions demonstrated that our system consistently outperformed several mainstream machine learning baselines, thereby supporting its effectiveness and potential generalisability across diverse clinical environments. In clinical practice, clinicians can use the MPSUR to predict the cause of unplanned reoperation based on a patient’s EMR. They can then integrate the high-confidence cause analysis results with their clinical expertise to develop a customised surgical plan for the patient. This study presents one of the first AI-based tools specifically designed for predicting the causes of unplanned reoperations using EMRs12 and provides a foundation for developing strategies to improve patient safety in surgical care. Methods Study design and participants This retrospective cohort study examined the causes of unplanned reoperations using data collected from 50 departments across 8 centres. The participating departments included general surgery, orthopaedics, cardiothoracic surgery, thoracic surgery, urology, neurosurgery, vascular surgery, plastic and cosmetic surgery, breast surgery, hepatic surgery, anaesthesiology, obstetrics and gynaecology, ophthalmology, otolaryngology, and stomatology. All participating institutions routinely recorded and reported unplanned reoperations as part of standard clinical practice. Inclusion criteria: All surgical patients in participating centres who experienced an unexpected return to the operating room for surgery. Exclusion criteria were patients who underwent reoperations unrelated to the initial procedure, trauma cases and transplantation cases. The dataset comprised 1781 cases from Peking Union Medical College Hospital (PUMCH) (2015–2024), 202 from People’s Hospital of Xingtai (2022–2024), 576 from Cangzhou Central Hospital (2022–2024), 224 from Zhangzhou Municipal Hospital (2022–2024), 34 from Beijing Cancer Hospital (2024), 42 from Beijing No.6 Hospital (2024), 47 from Beijing Puren Hospital (2024) and 16 from Beijing Longfu Hospital (2024). Data collection To ensure data quality, extraction was performed independently by two reviewers working in pairs, using admission records, preoperative discussions and anaesthesia assessment forms. Any discrepancies were resolved through consultation with a third reviewer. The indication for reoperation was determined by the primary surgeon and subsequently categorised by quality control personnel into three groups based on the underlying cause: bleeding, infection or other factors. Patient and public involvement Patients or the public were not involved in the design, or conduct, or reporting, or dissemination plans of our research. Feature description High-risk indicators and their clinical significance were adapted from the American College of Surgeons National Surgical Quality Improvement Program (ACS NSQIP) Surgical Risk Calculator (https://riskcalculator.facs.org/RiskCalculator/) and tailored to the objectives of this study. Consequently, the incorporated indicators comprised 14 baseline patient conditions at admission, details of the initial operation and a comprehensive list of all diagnoses and procedures recorded at discharge, culminating in 28 variables (table 1, feature description). Table 1Feature description, coding scheme and encoded structured features for the sample patientsFeature descriptionSeriesValue: age, sex, height, weight and BMIPrior to admission: emergency case, ascites, ventilator dependent, disseminated cancer, diabetes, hypertension requiring medication, congestive heart failure, dyspnoea, history of COPD, dialysis, acute renal failure, current smoker within 1 yearFirst operation: utilisation of robot-assisted surgery, ASA classFunctional status: this refers to the functional status of the patients on admission, specifically denoting whether they are fully independent in walking or not.TextFirst operation: diagnosis before the initial operation, name of first operation, department of the initial procedureDischarge: main procedure, all procedures, main diagnosis, other diagnosis Coding scheme Regarding sex feature, male is encoded as 1, female as 0 and missing values are encoded as −1.For continuous features such as height, weight and BMI, missing entries are imputed with a placeholder value of −1, while all valid values are retained without modification.The ASA classification is numerically encoded as follows: missing = −1, I=1, II=2, III=3, IV=4, V=5, IE=6, IIE=7, IIIE=8 and IVE=9. (E means emergency surgery)For binary features indicating ‘yes’ or ‘no’, a value of 1 denotes ‘yes’, 0 denotes ‘no’ and −1 is used to indicate missing data. Encoded structured features for the sample patients SexAgeASABinary variablesHeightWeightBMIFemale40IINo, Yes, No, No, No, No, No, No, No, No, No, No, No, No1635922.206Encoded structured features: (0, 40, 2, 0, 1, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 0, 163, 59, 22.206)Male88IIINo, Yes, No, NA, No, No, Yes, No, No, No, Yes, No, No, NA1756320.571Encoded structured features: (1, 88, 3, 0, 1, 0, −1, 0, 0, 1, 0, 0, 0, 1, 0, 0, −1, 175, 63, 20.571)ASA, American Society of Anesthesiologists; BMI, body mass index; COPD, chronic obstructive pulmonary disease. Feature description, coding scheme and encoded structured features for the sample patients ASA, American Society of Anesthesiologists; BMI, body mass index; COPD, chronic obstructive pulmonary disease. Our dataset comprises two distinct categories of cases. The first category includes patients who underwent two surgical procedures within the same hospital admission (n=2541). The second category includes patients who underwent two surgical procedures across two separate hospital admissions (n=381). For cases in which both surgeries occurred during the same admission, we excluded all discharge-related fields to avoid potential information leakage. Specifically, the removed features included the main procedure, all procedures, main diagnosis and other diagnoses, as these fields are typically finalised after the reoperation event. In contrast, for cases in which the two surgeries occurred in separate admissions, discharge-related information from the first admission was retained, as it was fully available prior to the subsequent surgery and therefore does not introduce temporal leakage. Overview of the MPSUR The clinical pathway of the MPSUR framework is illustrated in figure 1a. For patients undergoing unplanned reoperations, MPSUR first analyses the potential causes. The predicted results are then integrated with clinicians’ assessments to reach the final cause analysis, on which subsequent treatment plans are designed and implemented. (a) The clinical pathway of the MPSUR framework. (b) Overview of the MPSUR. The system consists of three stages: data preprocessing, feature extraction and AI model prediction. The MPSUR processes EMRs by categorising them into two distinct modalities, each of which can be used independently to predict the causes of unplanned reoperations, or jointly through a fused representation for integrated prediction. AI, artificial intelligence; EMR, electronic medical record; GCN, Graph Convolutional Network; LLM, large language model; MPSUR, Multi-modal Prediction System for Causes of Unplanned Reoperation; TFRNN, Time and Frequency Recurrent Neural Network. An overview of the MPSUR framework is illustrated in figure 1b. It consists of three stages: data preprocessing, feature extraction and AI model prediction. In the data preprocessing stage, the system processes EMRs by dividing them into two distinct modalities: structured series and text data. The series data consisted of 21 demographic and clinical features, including age, sex and height etc. The text data included seven categories of free-text fields such as surgical department, procedure name and diagnosis etc. For the feature extraction stage, the series data were processed using our designed encoding method and further reconstructed into graph-based features to better capture the inter-relationships among variables. The text data were first embedded using a large language model (LLM).13 14 In the AI model prediction stage, the graph-based features were processed using a Graph Convolutional Network to extract high-level semantic representations, which were used both for individual prediction of reoperation causes via a linear classification layer and for subsequent multimodal fusion. The text features were passed through a Time and Frequency Recurrent Neural Network (TFRNN),15 likewise enabling both standalone prediction and multimodal fusion. As the two data modalities vary in length and information density, their respective TFRNNs differ in terms of network depth and width. However, the final semantic representations were standardised to 128 dimensions. These representations are concatenated and passed through a final linear layer to predict the likely cause of unplanned reoperation. This modular design offers practical flexibility, enabling the model to function effectively even when one data modality is missing, thereby supporting use in real-world clinical settings where EMR completeness may vary. Coding and reconstructing for series data As shown in table 1 (coding scheme), we applied a consistent coding scheme across all 20 structured features. Missing data were uniformly encoded as −1. Sex was recorded as 1 for male and 0 for female. Continuous variables, including height, weight and body mass index (BMI), were retained in their original form. The American Society of Anesthesiologists (ASA)16 physical status classification was encoded numerically, with grades I to IVE represented by values ranging from 1 to 9. For the remaining 15 binary variables, responses were coded as 1 for ‘yes’ and 0 for ‘no’. The original values and their corresponding encoded representations are presented in table 1 (encoded structured features for the sample patients). To model the intrinsic relationships among the structured features, we constructed a graph representation where each of the 20 features corresponds to a node. The graph’s edges were defined based on the Pearson correlation matrix computed from the training data. An undirected edge was then established between any two nodes if the absolute value of their correlation coefficient was greater than a threshold of 0.5. Thus, each sample could be transformed into a structured graph with the same adjacency matrix and its corresponding node features. A sensitivity analysis was conducted in the online supplemental materials table 7 to assess the impact of different correlation thresholds (0.5, 0.4 and 0.3). The results showed that the model performance peaked at a threshold of 0.5, which represents a balance between graph sparsity and feature connectivity. Large language model The LLM played a crucial role in extracting features from unstructured clinical text. High-quality textual representations were essential for accurate downstream prediction of unplanned reoperation causes and effective multimodal feature fusion. In this study, we evaluated several mainstream language models, including BERT (Bidirectional Encoder Representations from Transformers),17 LLaMA (Large Language Model Meta AI) V.218 and LLaMA V.3.19 Experimental results demonstrated that BERT outperformed the others in this task, based on mean accuracy for predicting causes of unplanned reoperation, with performance of 62.03% (95% CI 60.54% to 63.50%) vs 59.42% (95% CI 57.83% to 61.01%) for LLaMA 2 and 59.66% (95% CI 57.96% to 61.35%) for LLaMA V.3. Detailed experimental results are provided in online supplemental table 1. Training of the MPSUR In the MPSUR framework, model training focuses on two structurally distinct time and TFRNNs, as well as three corresponding linear classifiers: one for the series modality, one for the text modality and one for the fused representation. The TFRNN designed for series data comprises a single convolutional layer followed by a Time-Frequency Block (TFBlock), reflecting the relative simplicity and lower dimensionality of structured inputs. By contrast, the TFRNN for text data incorporates two convolutional layers and three stacked TFBlocks to better capture the richer semantics and greater variability inherent in free-text clinical notes. During training, we primarily used the cross-entropy loss function to optimise the model. The specific form of the loss function can be expressed as: Loss = CE(YS,Y)+CE(YT,Y)+CE(YF,Y) where, Y denotes the label of the input data X , while YS , YT and YF represent the predictions of the series classifier, text classifier and fused representation classifier on the input data X , respectively. Statistical analysis of the dataset Statistical analyses were conducted as follows. Continuous variables were presented as mean±SD or median (IQR) based on the Shapiro-Wilk test for normality. Comparisons between the three outcome groups were performed using one-way analysis of variance for normally distributed variables or the Kruskal-Wallis H test for non-normally distributed variables. Categorical variables were summarised as counts and percentages, and differences between groups were evaluated using the χ2 test or Fisher’s exact test when expected frequencies were <5. All statistical tests were two tailed, and a p value<0.05 was considered statistically significant. Statistical analyses were performed using Python V.3.12.8. Detailed results are provided in online supplemental table 2. Results Experiment settings In this study, the datasets were categorised into internal and external. Data from institutions 1–3 were designated as the internal dataset (with a class distribution ratio of 2.8:2:1), whereas data from all other institutions were treated as the external dataset (with a class distribution ratio of 2.4:2.4:1). The internal dataset comprised 2559 cases, whereas the external dataset contained 363 cases. For the internal dataset, we split the data into training and testing sets using an 8:2 ratio and performed fivefold cross-validation with stratified sampling to maintain class distribution during both training and evaluation. The external dataset was used exclusively for testing to assess generalisability. The MPSUR model was trained for 1000 epochs using the Adam optimiser, with an initial learning rate of 0.001 and a batch size of 512. A stepwise learning rate decay strategy was applied, reducing the learning rate by a factor of 0.9 every 100 epochs. During training, the BERT encoder was kept frozen, and only the downstream task-specific modules were optimised. Given that the distribution of the three outcome categories showed no substantial imbalance, no class-weighting adjustments were applied. Model performance was primarily evaluated using classification accuracy, recall and F1 score, whereas the SD of accuracy across different splits was used to assess the robustness of each model. We computed 95% CIs) based on the results from five-fold cross-validation. Specifically, for each fold, we obtained the performance metrics on the test set, and then aggregated the results across the five folds to calculate the mean and SD. The 95% CI was subsequently estimated as the mean±1.96×SE across folds, providing an approximate measure of the variability in model performance. Baseline In this study, four widely used machine learning algorithms were employed as baselines: K-nearest neighbours20 with K values from 1 to 5; support vector machines (SVM)21 using linear,22 radial basis function23 and Sigmoid kernels24; decision trees25; naïve Bayes classifiers26; LogisticRegression (LR); and eXtreme Gradient Boosting (XGBoost). Evaluation of model accuracy in predicting causes of unplanned reoperation As summarised in table 2, the MPSUR framework was benchmarked against several widely used machine learning algorithms through fivefold cross-validation. The evaluation was conducted under three inference conditions: (1) single-modality inference using the series modality, (2) single-modality inference using the text modality and (3) full-modality inference combining all available data streams. This comprehensive design allowed for the assessment of the model accuracy in predicting the causes of unplanned reoperations across both complete and modality-limited scenarios, providing critical insights into its robustness and potential clinical utility. Table 2MPSUR was compared with other baseline models in terms of accuracy for cause prediction using fivefold cross-validation on both internal and external datasetsInternal datasetMethodFold 1Fold 2Fold 3Fold 4Fold 5Mean ↑SD ↓KNN (K=1)50.78%43.16%46.68%47.85%44.08%46.51%3.049%KNN (K=2)53.12%47.46%48.44%49.61%43.37%48.40%3.533%KNN (K=3)53.32%48.05%50.78%50.39%47.46%50.00%2.347%KNN (K=4)52.93%48.05%50.00%49.80%50.78%50.31%1.771%KNN (K=5)53.91%50.39%50.78%50.78%51.95%51.56%1.437%SVM (linear)49.22%48.83%49.41%47.66%43.44%47.71%2.483%SVM (RBF)51.95%47.66%47.27%43.55%49.90%48.07%3.148%SVM (Sigmoid)51.95%47.66%47.27%43.55%49.90%48.07%3.148%DecisionTree49.80%44.53%47.07%47.66%46.97%47.21%1.882%naive_bayes46.68%46.29%48.05%50.00%48.53%47.91%1.492%LR51.95%50.98%52.34%51.56%48.14%50.99%1.673%XGBoost51.08%53.91%55.47%54.99%55.19%54.13%1.803%Ours (series)54.69%53.41%50.00%50.00%55.66%52.75%2.636%Ours (text)60.16%60.23%58.09%60.16%59.96%59.72%0.917%Ours (all) 62.63% 60.34% 61.91% 61.72% 63.52% 62.03% 1.179% External dataset KNN (K=1)41.44%38.40%41.99%43.09%41.71%41.33%1.752%KNN (K=2)43.37%43.37%42.27%43.92%43.92%43.37%0.674%KNN (K=3)43.92%41.99%41.71%41.71%42.27%42.32%0.924%KNN (K=4)44.48%43.65%45.03%41.71%41.99%43.37%1.477%KNN (K=5)41.71%41.99%42.27%40.61%42.82%41.88%0.820%SVM (linear)39.23%47.51%42.27%36.46%48.07%42.71%5.078%SVM (RBF)41.44%41.44%41.44%41.44%41.44%41.44%0.000%SVM (Sigmoid)41.44%41.44%41.44%41.44%41.44%41.44%0.000%DecisionTree42.27%40.61%42.27%38.12%35.64%39.78%2.872%naive_bayes43.37%46.96%43.92%45.03%47.24%45.30%1.748%LR45.30%50.83%46.69%42.27%53.87%47.79%4.585%XGBoost47.79%45.86%43.65%45.03%48.34%46.13%1.941%Ours (series)45.30%45.30%46.13%48.63%44.75%46.02%1.539%Ours (text)55.25%54.97%56.63%53.31%56.08%55.25%1.268%Ours (all) 57.80% 54.35% 57.73% 55.52% 56.63% 56.41% 1.479%KNN, K-nearest neighbours; LR, LogisticRegression; RBF, radial basis function; SVM, support vector machines; XGBoost, eXtreme Gradient Boosting. MPSUR was compared with other baseline models in terms of accuracy for cause prediction using fivefold cross-validation on both internal and external datasets KNN, K-nearest neighbours; LR, LogisticRegression; RBF, radial basis function; SVM, support vector machines; XGBoost, eXtreme Gradient Boosting. On the internal test dataset, the MPSUR achieved a mean accuracy of 62.03% (95% CI 60.54% to 63.50%) for cause prediction using full-modality input, significantly outperforming the best-performing baseline—XGBoost—by approximately 8% (p<0.001). On the external dataset, the MPSUR attained a mean accuracy for cause prediction of 56.41% (95% CI 55.57% to 58.24%), exceeding the best-performing baseline (LR) by roughly 8.6% (p<0.01), thereby reinforcing the effectiveness of the proposed approach. When using only the text modality, the mean accuracy of the MPSUR for cause prediction decreased minimally—by 2.31% on the internal dataset and 1.16% on the external dataset—yet remained substantially superior to all baseline models, including those using full-modality input. Inference using the series modality showed a more pronounced performance reduction; however, accuracy remained higher than that achieved by most benchmark algorithms. These findings underscore the adaptability and strength of the MPSUR in real-world settings where data from certain modalities may be missing. Moreover, the decrease in mean accuracy of the MPSUR from internal to external datasets was approximately 7%, compared with a larger decline of approximately 8% observed in the best-performing baseline (XGBoost). This decrease likely reflects differences in patient populations and institutional protocols across hospitals. Feature distribution analysis revealed notable differences in class distribution between the internal and external datasets. The SD of accuracy for cause prediction across cross-validation folds was <2% for both internal and external datasets, indicating high stability and robustness of the model across different datasets and data-splitting strategies. In order to comprehensively evaluate the performance of MPSUR, recall and F1 score were additionally assessed, as presented in online supplemental tables 3 and 4. On the internal dataset, MPSUR achieved a mean recall of 0.5443 (95% CI 0.5262 to 0.5624) and a mean F1 score of 0.5277 (95% CI 0.5092 to 0.5462), markedly outperforming the best-performing machine learning model, XGBoost, which yielded 0.4757 (95% CI 0.4612 to 0.4902) and 0.4763 (95% CI 0.4578 to 0.4948), respectively. On the external dataset, MPSUR maintained superior performance with a mean recall of 0.5050 (95% CI 0.4867 to 0.5233) and a mean F1 score of 0.4706 (95% CI 0.4481 to 0.4931), substantially exceeding the corresponding metrics obtained by LR, which were 0.4262 (95% CI 0.3807 to 0.4717) and 0.4357 (95% CI 0.3979 to 0.4735), respectively. These experimental findings further demonstrate the substantial advantages of MPSUR over other approaches. Multidepartmental clinical reader study To evaluate the clinical applicability of our proposed MPSUR framework, we compared its accuracy for cause prediction with that of clinicians from four departments at the PUMCH on a set of 90 cases requiring unplanned reoperation. Participating clinicians included three surgeons from general surgery (mean experience: 14 years), three from thoracic surgery (mean experience: 16 years), one from cardiac surgery (mean experience: 13 years) and three from emergency surgery (mean experience: 10 years). Cases were presented in randomised order, with all identifying information removed. Clinicians were provided with the same input features used by the MPSUR and asked to predict the most likely cause of reoperation. Time limitations were not imposed. Inter-rater reliability was assessed using Fleiss’ kappa, achieving K=0.040 (95% CI −0.02 to 0.10). As illustrated in figure 2, MPSUR demonstrated a statistically significant advantage (mean accuracy=62.68%, 95% CI 60.36% to 65.01%) over clinicians across all departments in predicting the causes of unplanned reoperations with mean accuracy differences of 22.68% (95% CI 20.06% to 25.30%) for general surgery, 23.57% (95% CI 20.88% to 26.26%) for thoracic surgery, 14.50% (95% CI 12.50% to 17.3%) for cardiac surgery and 13.39% (95% CI 10.46% to 16.31%) for emergency surgery, demonstrating consistent superiority across all departments. This highlights the potential of the MPSUR as a promising clinically applicable tool for enhancing the accuracy of decision-making related to unplanned surgical interventions. Clinical reader study comparing the accuracy of the MPSUR with that of clinicians. The study involved four surgical departments at the institution: general surgery (GS Dept.), thoracic surgery (TS Dept.), cardiac surgery (CS Dept.) and emergency surgery (ES Dept.). Results show that the MPSUR consistently outperformed human experts across all departments, demonstrating its potential to support clinical decision-making. MPSUR, Multi-modal Prediction System for Causes of Unplanned Reoperation. Feature ablation study We conducted a feature ablation study to assess the relative contributions of different feature categories to the accuracy of MPSUR in predicting causes of unplanned reoperations. As summarised in table 3, all series features were grouped into a single category labelled ‘series’ owing to their relatively low dimensionality. The seven text-based features (text 1 to text 7) corresponded to department, diagnosis before surgery, procedure, main diagnosis, other diagnoses, main surgery and surgery. Considering that the main diagnosis, other diagnoses, main surgery and surgery (text 4 to text 7) represent discharge information and that the available data for these features were limited (n=381), we grouped them into a single category for the ablation study. The ‘w/o’ indicates that the corresponding feature was removed during both training and inference, whereas ‘w/’ denotes that only the corresponding feature was used during both training and inference. Table 3Feature ablation study conducted to evaluate the individual contribution of each input variable to the model’s predictive performanceIntelnal datasetMethodFold 1Fold 2Fold 3Fold 4Fold 5AverageP valueOurs (series)54.69%53.41%50.00%50.00%55.66%52.75%–w/ series51.85%53.22%47.56%48.34%53.04%50.80%0.0002Ours (text)60.16%60.23%58.09%60.16%59.96%59.72%–w/o series58.67%59.57%57.89%58.77%59.26%58.83%0.0018Ours (all)62.63%60.34%61.91%61.72%63.52%62.03%–w/o text 159.06%60.62%58.48%59.57%58.79%59.30%0.0129w/o text 260.04%60.35%59.18%59.65%61.21%60.09%0.0018w/o text 359.65%59.65%58.70%57.50%59.38%58.98%0.0025w/o text{ 4, 5, 6, 7}61.04%59.96%61.52%61.52%60.16%60.84%0.0900 External dataset MethodFold 1Fold 2Fold 3Fold 4Fold 5AverageP valueOurs (series)45.30%45.30%46.13%48.63%44.75%46.02%–w/ series45.03%44.75%45.86%45.30%41.37%44.46%< 0.0001Ours (text)55.25%54.97%56.63%53.31%56.08%55.25%–w/o series53.59%53.31%54.42%53.04%53.87%53.65%0.0032Ours (all)57.80%54.35%57.73%55.52%56.63%56.41%–w/o text 154.14%53.31%54.42%53.48%54.14%53.90%0.0020w/o text 254.63%55.31%56.08%55.25%55.25%55.30%0.1300w/o text 354.97%52.76%51.66%52.21%53.31%52.98%0.0018w/o text{ 4, 5, 6, 7}55.21%56.63%56.61%55.04%54.63%55.62%0.1300 Feature ablation study conducted to evaluate the individual contribution of each input variable to the model’s predictive performance As presented in table 3, we compared the performance of our method under two settings: full-modality training followed by single-modality inference (ours (series) and ours (text)), and both training and inference using only a single modality (table 3, w/ series and w/o series). The results indicated that models trained with full-modality inputs achieved better performance, even when tested using a single modality. This suggests that full-modality training enables cross-modal knowledge transfer, allowing each modality to benefit from the complementary information provided by others, thereby enhancing predictive accuracy. Notably, the removal of text 4 to text 7 had the smallest impact on performance. On the internal dataset, removing text 4 to text 7 led to a decrease in accuracy of 1.18% (95% CI −0.05% to 2.41%), and on the external dataset, a decrease of 0.78% (95% CI −1.17% to 2.73%). In contrast, removing text 3 (procedure) resulted in the most pronounced performance decline, with accuracy decreasing by 3.05% (95% CI 1.51% to 4.60%) on the internal dataset and 3.42% (95% CI 1.59% to 5.26%) on the external dataset. Statistical comparisons using p values between each ablated model and the full-modality model (ours (all)) showed that the models with text 2 removed, as well as text 4 to text 7 removed, had other models’results with p values<0.05, indicating statistically significant differences. These findings highlight the critical contribution of these features—specifically the series data, department and procedure—to the predictive performance of MPSUR. To further investigate the importance of the 21 structured series features in predicting the causes of unplanned reoperation, we conducted a standalone analysis using two classical machine learning models: SVM (linear) and a decision tree. Each feature was individually removed from the training and testing sets to observe its impact on model performance. Given the high correlation between height, weight and BMI, we retained only BMI and excluded the other two variables. As summarised in online supplemental tables 5 and 6, the top five most influential features for the SVM were BMI, utilisation of robot-assisted surgery, ASA class, age and dyspnoea. For the decision tree, the most influential features were sex, utilisation of robot-assisted surgery, age, BMI and ASA class. Collectively, four shared features—BMI, utilisation of robot-assisted surgery, ASA class and age—emerged as key predictors across both models. These shared series features represent the primary discriminative factors for identifying the causes of unplanned reoperations. Discussion This study demonstrates that the MPSUR, an AI-based multimodal prediction system, significantly outperforms traditional machine learning approaches in predicting the causes of unplanned reoperations. The system achieved 62.03% accuracy in internal validation and maintained reasonable performance (56.41%) in external validation, representing a substantial improvement over existing approaches. In the clinical reader study, MPSUR also outperformed experienced clinicians, achieving a mean accuracy=60.27% (95% CI 58.27% to 62.26%), highlighting its potential as a clinical decision-support tool. We acknowledge that the clinical reader study is necessarily more limited in scope compared with the technical validation; however, it still demonstrates that MPSUR provides meaningful guidance beyond routine clinician assessment. Its ability to function with incomplete data modalities enhances practical applicability in real-world settings, where EMR completeness varies. Integration into existing hospital information systems (HIS) could enable real-time risk assessment and targeted interventions. Specifically, before an unplanned reoperation, clinicians can combine high-confidence predictions from the MPSUR with their expertise to tailor surgical plans for patients. We rigorously followed the ACS NSQIP Surgical Risk Calculator framework to select EMR data features. This standardised approach is applicable across EMRs in different languages, ensuring that the MPSUR can be deployed in countries or regions with diverse linguistic environments. Feature ablation analyses underscored the critical importance of certain text-based features—specifically, procedure details, main diagnosis and surgical information—for accurate prediction. In addition, structured series features such as BMI, ASA class, age and use of robot-assisted surgery were consistently identified as key predictors across different machine learning models. These findings reinforce the clinical relevance of incorporating both structured and unstructured data for comprehensive risk assessment. However, the 7% performance decrease on external validation highlights important considerations for implementation. This decrease likely reflects differences in patient populations, coding practices and institutional protocols between hospitals. Local adaptation and continuous model updating would be necessary for optimal performance across different healthcare systems. MPSUR employs multimodal training to enable cross-modal knowledge transfer, mitigating the impact of missing or low-quality data. Careful preprocessing of input data is required to ensure optimal performance. Given the complexity of the model and the integration of diverse data types, clinical use requires users with data processing skills and the ability to interpret probabilistic outputs. Therefore, adequate training is essential for the successful implementation of MPSUR in practice. Future studies should focus on expanding the applicability of MPSUR across diverse clinical settings and data sources, leveraging EMRs to predict the risk of unplanned reoperations in newly admitted patients and exploring model-updating strategies to adapt to evolving clinical practices and data characteristics. In addition, statistical analysis of unplanned reoperations should consider the classification and management of surgical procedures, as different disease groups may exhibit varying rates and underlying causes. Regular monitoring and structured data collection—such as monthly analyses within HIS—can help identify potential systemic or technical factors in a timely manner, enhancing the utility of predictive tools such as MPSUR in clinical practice. Overall, MPSUR represents a promising tool for predicting the causes of unplanned reoperations, with the potential to improve surgical care quality and patient outcomes.