Continuous Multimodal AI with Wearable Vital Signs Predicts Postoperative Complications in the General Ward

Surgery is inherently associated with complications, making early detection the cornerstone of timely intervention and improved outcomes. Artificial intelligence (AI) has been shown to predict severe events such as sepsis and mortality after surgery within intensive care units (ICUs). However, most complications occur on general wards (GW), where staffing and technical monitoring constraints impede effective real-time detection of complications. Here, we present a real-time, multimodal AI-based complication prediction system for surgical GWs, combining routine clinical data with continuous high-resolution vital signs derived from telemetric photopletysmography (PPG) sensors into digital patient representations. A total of 1,285 patients undergoing esophageal, gastric, liver, pancreatic, and colorectal surgery were prospectively enrolled. Baseline patient characteristics, intraoperative data, ICU parameters, and GW data, including 270,603 hours of recorded telemetric vital signs, were used to detect postoperative intra-abdominal infections. We demonstrate a high median area under the receiver operator characteristic (AUROC) of 0.90 (0.89-0.91) for the detection of surgery-related infections. Complications could be predicted 9 hours in advance with only a minor reduction of the AUROC: 0.89 (0.88-0.89). Continuous wearable data increased the AUROC by 8% and the Area Under the Precision-Recall Curve (AUPRC) by 109%, outperforming other modalities in our ablation experiments. Further development into an AI-based alarm system outperforms traditional early warning scores. These findings highlight the potential of high-dimensional, multimodal, real-time risk stratification to support earlier detection of adverse events in surgical patients. Our results reveal continuous monitoring, using minimally intrusive vital signs, as a key component of an intelligent, data-driven smart ward.