Adaptive neuro-fuzzy inference systems with k-fold cross-validation for energy expenditure predictions based on heart rate

• We present a practical approach based on neuro-fuzzy systems to estimate energy expenditure using heart rate monitoring.
• The proposed approach improves the standard Flex-HR method in that it does not require individual calibration.
• The proposed approach treats the uncertainty in human physiological systems and in various workplaces by using fuzzy logic.

This paper presents a new model based on adaptive neuro-fuzzy inference systems (ANFIS) to predict oxygen consumption (V˙O2) from easily measured variables. The ANFIS prediction model consists of three ANFIS modules for estimating the Flex–HR parameters. Each module was developed based on clustering a training set of data samples relevant to that module and then the ANFIS prediction model was tested against a validation data set. Fifty-eight participants performed the Meyer and Flenghi step-test, during which heart rate (HR) and V˙O2 were measured. Results indicated no significant difference between observed and estimated Flex–HR parameters and between measured and estimated V˙O2 in the overall HR range, and separately in different HR ranges. The ANFIS prediction model (MAE = 3 ml kg−1 min−1) demonstrated better performance than Rennie et al.'s (MAE = 7 ml kg−1 min−1) and Keytel et al.'s (MAE = 6 ml kg−1 min−1) models, and comparable performance with the standard Flex–HR method (MAE = 2.3 ml kg−1 min−1) throughout the HR range. The ANFIS model thus provides practitioners with a practical, cost- and time-efficient method for V˙O2 estimation without the need for individual calibration.