OA01.4. Multimodal Early Warning System for Postoperative Complications After Esophagectomy: Integrating Patient-Reported Outcomes, Laboratory Trajectories, and Clinical Text Mining
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.
