Optimization of Operating Room Monitor Alarm Sounds Based on Intelligent Audio Processing

Background: Operating rooms (ORs) are high-noise environments where frequent monitoring of alarms contributes to alarm fatigue, reducing healthcare providers’ response efficiency and potentially compromising patient safety. Traditional alarm systems rely on preset thresholds without intelligent optimization, leading to frequent false alarms and cognitive overload for surgical teams. Recent advances in artificial intelligence (AI) and psychoacoustics offer new opportunities to enhance alarm perception and improve response times. Objective: This study aims to optimize monitor alarm sounds in the OR by integrating deep learning-based alarm classification (CNN + LSTM) with psychoacoustic modeling to reduce auditory fatigue and improve response efficiency. Methods: A total of 35 OR healthcare professionals (15 anesthesiologists, 12 surgeons, 8 nurses) were recruited from a tertiary hospital using purposive sampling. Participants were randomly assigned to a control group (standard alarms) or an experimental group (optimized alarms). Alarm Sound Processing & Classification: Real-world OR alarm sounds (ECG, SpO？, ventilators, infusion pumps) were recorded and analyzed. Psychoacoustic parameters (loudness, sharpness, unpleasantness index) were computed using Zwicker’s model, Aures formula, and Glasberg & Moore approach. A CNN+LSTM model was trained using Mel-Frequency Cepstral Coefficients (MFCCs) to classify alarms into critical, non-critical, and false alarms. Data were split into training (70%), validation (15%), and testing (15%) sets. Experimental Design & Evaluation: Participants completed two OR scenarios: one using traditional alarms and the other using optimized alarms. Response times, alarm recognition accuracy, and subjective fatigue levels were measured. Fatigue was assessed using NASA Task Load Index (NASA-TLX), Mental Fatigue Scale (MFS), and a Likert-based fatigue rating (1 - 7 scale). Post-hoc power analysis confirmed that the study was adequately powered (power = 0.82). Results: CNN+LSTM Model Performance: Achieved 92.4% classification accuracy, with false alarms reduced by 37%. Alarm Response Time: Improved by 25% (2.4 s → 1.8 s, p < 0.01) in the experimental group. False Alarm Reduction: Participants responded to only 10% of false alarms with optimized sounds, compared to 70% with traditional alarms. Fatigue & Cognitive Load: NASA-TLX workload scores dropped from 72 to 58. Fatigue ratings decreased from 4.1 to 2.7 (Likert scale, p < 0.05). 80% of participants