OA01.4. Multimodal Early Warning System for Postoperative Complications After Esophagectomy: Integrating Patient-Reported Outcomes, Laboratory Trajectories, and Clinical Text Mining
Aug 2026· Diseases of the esophagus· Vol 39· 0 citations
TL;DR
A multimodal early warning system integrating PROs, laboratory trajectories, and large language model-extracted clinical text features to enable real-time, automated complication surveillance substantially outperforms single-modality approaches for detecting postoperative complications after esophagectomy.
Abstract
Esophageal Cancer: Surgical Treatment of Esophageal Cancer – early outcomes and complications
Postoperative complications after esophagectomy remain a major source of morbidity, yet current surveillance relies predominantly on clinician-initiated assessments and structured database fields that systematically underdetect complications. We aimed to develop and internally validate a multimodal early warning system (MEWS-Eso) integrating patient-reported outcomes (PROs), laboratory trajectories, and large language model (LLM)-extracted clinical text features to enable real-time, automated complication surveillance.
This prospective cohort study enrolled 323 consecutive patients undergoing esophagectomy for esophageal cancer at a single tertiary center (2019–2024). PROs were collected using the MD Anderson Symptom Inventory (MDASI) at 10 timepoints from preoperative baseline through 6 months postoperatively (completion rate 91.3%). Laboratory values (complete blood count, C-reactive protein, procalcitonin, albumin) were extracted at 7 perioperative timepoints. Free-text clinical notes (nursing records, physician progress notes, operative reports; median 3,247 words/patient for POD0–POD7) were processed using GPT-4o for structured information extraction and semantic embedding generation. NLP was independently applied to identify complications missed by structured fields. Six models of incrementally increasing modality were compared using 5-fold cross-validation with 100 bootstrap iterations: M1 (PRO-only), M2 (PRO + clinical baseline), M3 (PRO + laboratory), M4 (PRO + text embeddings), M5 (PRO + laboratory + text), and M6 (full multimodal). The primary endpoint was postoperative complications within 90 days under both standard and NLP-augmented outcome definitions.
NLP text mining identified 72 complications missed by structured fields, most notably anastomotic stricture (43 NLP-detected vs. 3 structured-field recorded; p < 0.001). Under the NLP-augmented definition, complication prevalence increased from 52.3% to 61.0%. AUROC improved progressively with modality addition: M1 (PRO-only) 0.658 [95% CI 0.601–0.715], M3 (PRO + laboratory) 0.761 [0.712–0.810], M5 (PRO + laboratory + text) 0.824 [0.779–0.869], and M6 (full multimodal) 0.847 [0.805–0.889]. The CRP trajectory (POD1–POD5 slope) was the single strongest laboratory predictor (OR 2.41, 95% CI 1.78–3.26). LLM-extracted text features contributed an incremental AUROC gain of +0.063 beyond PRO + laboratory. At the RED-alert threshold (probability ≥0.70), MEWS-Eso achieved sensitivity 72.8%, specificity 83.5%, and positive predictive value 78.3%, with a median early warning lead time of 4.2 days. Removing the PRO module caused the largest performance drop (ΔAUROC = −0.089), followed by laboratory (−0.074) and text (−0.063).
A multimodal early warning system integrating PROs, laboratory trajectories, and LLM-processed clinical text substantially outperforms single-modality approaches for detecting postoperative complications after esophagectomy. PROs provide the most irreplaceable modality, while NLP-based text mining both corrects systematic outcome misclassification and contributes independent predictive signals. This framework demonstrates the feasibility of automated, patient-centered multimodal surveillance in surgical oncology.
Esophageal Cancer: Surgical Treatment of Esophageal Cancer – early outcomes and complications
Postoperative complications following oesophagectomy occur in 25–60% of patients, yet traditional patient-reported outcome (PRO) assessment relies on standardised questionnaires with limited reliability, particularly among elderly patients with low health literacy. We aimed to develop and externally validate a multimodal artificial intelligence (AI) system that automates PRO assessment from natural patient conversations and provides early warning of major postoperative complications.
We conducted a prospective observational study at a tertiary cancer center, enrolling consecutive patients undergoing esophagectomy.The development cohort (Centre 1, n=196) and the temporally independent external validation cohort (Centre 2, n=103) were recruited sequentially. Using the validated Patient Symptom Assessment for Oesophageal Cancer (PSA-ESO) instrument, we collected trimodal recordings (video, audio, text) at up to 25 timepoints per patient, yielding 6,813 evaluable assessments. We fine-tuned the Qwen2.5-Omni-7B multimodal large language model with LoRA adaptation for two tasks: automated PRO symptom grading (Task A) and Temporal Transformer-based early warning of Clavien-Dindo grade II or higher complications (Task B). In a prospective implementation substudy (n=62), we evaluated the clinical impact of real-time alerts on time-to-intervention.
In the external validation cohort, the trimodal PRO assessment achieved a weighted kappa of 0.801 (95% CI 0.73–0.87) and ICC of 0.858 against expert consensus, significantly outperforming audio-text bimodal (kappa 0.754, p<0.0001) and text-only (kappa 0.689, p<0.0001) configurations. The early warning model achieved an AUROC of 0.873 (95% CI 0.82–0.93) with a mean detection lead time of 28.7 hours before clinical diagnosis. The system detected 76.1% of symptom under-reporting cases. In the implementation substudy, real-time alerts reduced median time-to-intervention from 14.2 hours to 6.8 hours (p=0.003) and were associated with shorter ICU stays (3.1 vs 5.4 days, p=0.028).
An end-to-end multimodal AI system can accurately automate PRO assessment from natural patient conversations and provide clinically meaningful early warning for postoperative complications, with external validation confirming generalisability across cohorts. This conversation-based approach represents a paradigm shift from questionnaire-based PRO evaluation in surgical oncology, with particular relevance for populations with limited health literacy.
Si-miao Lu, Yi Zhu, Yong-tao Han et al.· Diseases of the esophagus· 0 citations
Esophageal Cancer: Surgical Treatment of Esophageal Cancer – early outcomes and complications
Anastomotic leak (AL) is one of the most severe complications after esophagectomy. Early identification of high-risk patients may help optimize perioperative management and improve outcomes. Machine learning (ML) offers promising tools for clinical prediction, yet studies applying ML to predict AL risk after esophagectomy remain scarce. This study aimed to develop and compare ML-based models using single-center real-world data to identify patients at high risk for AL.
Consecutive patients undergoing esophagectomy from April 2019 to April 2023 were retrospectively enrolled. The primary outcome was AL within 30 postoperative days. Thirty-seven potential predictors covering demographics, comorbidities, laboratory values, perioperative factors, and tumor characteristics were collected. Data were split 7:3 into training and test sets. Missing continuous variables were handled by multiple imputation, categorical variables by mode imputation. All features were standardized. Univariate ANOVA (P<0.10) followed by LASSO logistic regression was used for feature selection. Eight ML models were trained and tuned via 5-fold cross-validation with grid search. Performance was evaluated using AUC, accuracy, F1 score, calibration and decision curves. SHAP analysis was performed for interpretability. All analyses were performed in Python with scikit-learn, XGBoost, and SHAP packages.
502 patients who underwent esophagectomy were included, with an overall AL rate of 19.3% (97/502). Following data preprocessing and multiple imputation, univariate analysis identified 15 candidate predictors, which were refined to 15 model-level features using LASSO logistic regression. Eight machine learning models were trained. The Stochastic Gradient Boosting Tree (SGBT) achieved the highest AUC in both the training set (0.880) and the test set (0.683, Figure A). Model calibration demonstrated good agreement between predicted and observed outcomes (Figure B). The SGBT model yielded the highest recall (0.456, Figure C), and was therefore selected as the final predictive model. SHAP analysis revealed that radiotherapy, hemoglobin difference, intraoperative bleeding volume, and lymphocyte count were the most influential features driving individual risk predictions (mean absolute SHAP values shown in Figure D; full summary in Figure E). An integrated risk map visualizing the contribution of each key feature is provided in Figure F.
This single-center study developed ML models for predicting AL after esophagectomy. The SGBT model demonstrated acceptable discriminative ability and identified key risk factors including radiotherapy, hemoglobin drop, bleeding, and lymphocyte count. Despite retrospective limitations, these findings provide a data-driven foundation for risk stratification. Future multicenter validation and prospective studies are warranted to verify clinical utility.
Chi Zhang, Yongde Liao· Diseases of the esophagus· 0 citations
Esophageal Cancer: Surgical Treatment of Esophageal Cancer – early outcomes and complications
Persistent moderate-to-severe symptoms after minimally invasive esophagectomy (MIE) remain poorly characterized. While patient-reported outcomes (PROs) are increasingly recognized as important endpoints in surgical oncology, the prevalence, risk factors, optimal intervention timing, and prognostic impact of persistent postoperative symptom burden in esophageal squamous cell carcinoma (ESCC) patients have not been systematically evaluated. Understanding these factors is essential for developing evidence-based, timely symptom management strategies to improve postoperative recovery and quality of life.
We retrospectively analyzed 345 ESCC patients who underwent MIE at a single tertiary center. Symptoms were assessed using the MDASI, EORTC QLQ-C30, and QLQ-OES18 at baseline through 6 months postoperatively. Persistent moderate-to-severe symptoms at postoperative month 3 were defined by composite criteria (meeting any one): MDASI ≥3 core symptoms scoring ≥4; QLQ-C30 ≥2 functional domains below 66.7; or OES18 ≥2 symptom domains above 33.3. Independent risk factors were identified by multivariable logistic regression. Symptom trajectory analysis identified the critical divergence timepoint. Kaplan–Meier analysis, log-rank tests, multivariable Cox regression, and propensity score matching (1:1, 81 pairs) evaluated associations between symptom persistence and overall survival (OS) and disease-free survival (DFS).
At postoperative month 3, 89 patients (25.8%) met criteria for persistent moderate-to-severe symptoms. Independent risk factors were anastomotic leak (OR=7.92, P<0.001), adjuvant therapy (OR=3.05, P<0.001), and smoking history (OR=2.58, P=0.029); male sex was protective (OR=0.34, P=0.013). Predictive model AUC was 0.734. Symptom trajectories diverged at postoperative day 14. Five-year OS (63.5% vs 66.1%, HR=1.16, P=0.478) and DFS (72.9% vs 73.8%, HR=0.87, P=0.554) showed no significant differences. After propensity score matching (81 pairs), OS remained non-significant (HR=0.94, P=0.796), while DFS favored the persistent group (HR=0.55, P=0.029), potentially mediated by higher adjuvant therapy exposure.
Approximately one-quarter of post-MIE ESCC patients experience persistent moderate-to-severe symptoms at 3 months, primarily driven by anastomotic leak and adjuvant therapy. Symptom persistence does not adversely affect overall survival; its association with improved DFS likely reflects the tumor-control benefit of adjuvant therapy. Postoperative day 14 represents the critical symptom trajectory divergence point, providing an optimal window for initiating targeted PRO-based intervention.
Si-miao Lu, Yi Zhu, Yong-tao Han et al.· Diseases of the esophagus· 0 citations
Background Early postoperative pneumonia (POP) is a common and serious complication after brain tumor surgery, but early recognition is difficult because postoperative neurological dysfunction and respiratory symptoms are often non-specific. Existing models are mostly retrospective, not designed for neurosurgical patients, and rarely prospectively validated across centers. We aimed to develop an interpretable model for early POP risk stratification. Methods We used routine perioperative data from 1,856 patients undergoing brain tumor surgery at multiple centers in China between 2022 and 2025. Ten machine learning algorithms were compared. From 41 candidate variables, 11 predictors were selected using correlation analysis and LASSO. The final locked model was prospectively tested in one internal temporal cohort and three external cohorts. Performance was assessed by AUC, calibration, and decision curve analysis. Interpretability was evaluated using SHAP, a nomogram, and a web calculator. Results Logistic regression showed the best overall performance, with an AUC of 0.897 (95% CI, 0.842–0.952) in the internal cohort and a mean AUC of 0.876 ± 0.044 across the three external cohorts. Key predictors included chronic lung disease (CLD), diabetes mellitus (DM), body mass index (BMI), admission Karnofsky Performance Status (KPS), and preoperative albumin (Alb) and glucose (Glu). Conclusion This interpretable 11-variable model enables early POP risk stratification after brain tumor surgery and may support timely preventive intervention in neurosurgical care.
H. Mao, Fengchun Mu, Xinyu Wang et al.· Frontiers in Cellular and In...· 0 citations
Postoperative CRP trajectory patterns are significantly associated with complication severity after esophagectomy, and the High-Persistent trajectory identifies a high-risk subgroup, and the D3/D1 ratio enables early risk stratification by POD 3.
Si-miao Lu, Yi Zhu, Yong-tao Han et al.· Diseases of the esophagus· 0 citations
Introduction Early recurrence (typically within 2 years) following curative liver resection remains the primary obstacle to long-term survival in patients with hepatocellular carcinoma (HCC). Traditional linear staging systems frequently fail to capture the non-linear clinical and biological interactions driving early relapse. This study aimed to develop and validate a methodologically rigorous machine learning framework for precision post-hepatectomy risk stratification, and to deploy an interactive clinical tool. Patients and Methods In this large-scale retrospective cohort study, we analyzed 3084 HCC patients who underwent curative resection. To strictly prevent data leakage and ensure generalizability, the cohort was partitioned into training (60%), validation (20%), and independent test (20%) sets prior to any preprocessing. Six ML algorithms were evaluated, and SHapley Additive exPlanations (SHAP) were employed to decode model transparency. A web-based calculator was subsequently deployed for clinical use. Results XGBoost emerged as the optimal model, achieving an area under the curve (AUC) of 0.891 (95% CI: 0.865–0.915) on the independent test set, significantly outperforming traditional logistic regression (P < 0.0001). Using a rigorously optimized probability threshold of 0.310, the model yielded a sensitivity of 87.6% (95% CI: 83.5%–91.4%) and a specificity of 70.3% (95% CI: 65.4%–75.2%), with all confidence intervals derived from 2,000 bootstrap resamples. SHAP analysis identified tumor capsule integrity, neutrophil-to-eosinophil ratio (NER), and alpha-fetoprotein (AFP) as the most critical predictors of recurrence. Conclusion This methodologically rigorous ML framework provides a highly sensitive, data-driven tool for individual risk stratification. By identifying high-risk patients, the web-based calculator can guide clinicians in implementing intensified postoperative surveillance and selecting candidates for targeted adjuvant interventions, ultimately improving oncological outcomes in HCC.
Xuanyi Zhu, Yi-Teng Zhao, Mu-Yan Li et al.· Journal of Hepatocellular Ca...· 0 citations
Related blog posts
MIT News · Artificial Intelligence· news.mit.eduAug 27, 2026
A new machine-learning framework aims to improve the success rate of computational protein design while moving away from results that reproduce sequences found in nature.