Aprimoramento de modelo de inteligência artificial para estimativa de peso corporal de pacientes acamados por fotos de smartphones

Abstract

Reference: Enhancement of an artificial intelligence model for body weight estimation in bedridden patients using smartphone images Introduction: Estimating body weight in bedridden patients remains a frequent challenge in hospital practice when direct measurement is not feasible. Conventional indirect methods present operational limitations and substantial variability in accuracy, whereas artificial intelligence–based approaches have emerged as promising alternatives. Objectives: To develop and improve an artificial intelligence model for body weight estimation from smartphone-acquired images obtained in a real-world hospital environment, aiming to achieve a proportion of estimates within 10% of measured body weight (P10) greater than 70%. Methods: This cross-sectional analytical study was conducted in five public hospitals in the Federal District of Brazil and included 300 hospitalized adults and approximately 1,300 body images. After direct measurement of body weight and height, participants were photographed in a hospital bed to simulate the intended application scenario of the technology. The model employed body segmentation, three-dimensional reconstruction, extraction of anthropometric attributes, and regression techniques to estimate body weight. Performance was evaluated using the proportion of estimates within 10% of measured body weight (P10), Mean Absolute Error (MAE), coefficient of determination (R²), and Intraclass Correlation Coefficient (ICC). Results were compared with those of a previously developed pilot model. Results: The model achieved a P10 of 78.7%, indicating that 78.7% of weight estimates were within 10% of the measured body weight. Compared with the pilot model, the Mean Absolute Error decreased from 13.16 kg to 5.24 kg, corresponding to a mean paired improvement of 7.93 kg. The coefficient of determination increased from 0.388 to 0.873, and agreement between estimated and measured values was high (ICC = 0.93). Conclusions: The artificial intelligence model demonstrated potential for body weight estimation in a real-world hospital setting using smartphone images. The findings indicate substantial improvement compared with the pilot model and performance consistent with contemporary criteria for clinical acceptability, supporting its potential use as a tool to assist multidisciplinary clinical practice when direct body weight measurement is unavailable.