Real-time prediction of indoor thermal comfort has great potential in the development of a thermal comfort-driven control method for heating, ventilation, and air conditioning (HVAC) systems and the improvement of the indoor environment. This study measured physiological parameters (including electrocardiogram (ECG) signal, electroencephalogram (EEG) signal, and skin temperature (ST)) of 10 male and 10 female volunteers under three temperature conditions (22 °C, 25 °C, and 28 °C). A setup database containing 3600 sets of data was used to develop the thermal comfort prediction model with three machine learning algorithms: random forest (RF), back-propagation neural network (BP), and convolutional neural network (CNN). Significant differences were found in both the values and the impact indices of physiological indicators between males and females (p < 0.05). The RF algorithm had a significant advantage with an overall accuracy of 89.3 %. If only one factor is employed for machine learning training, using the RF algorithm and training with the ST dataset is the most appropriate method, with an accuracy of 81.8 %. The addition of all datasets can increase the modeling accuracy to 95.6 %. Taking into account the sample imbalance, the SMOTETomek method was applied. After applying the SMOTETomek method to balance the data, the model accuracy reached at 97.3 %. The comprehensive results showed that thermal comfort prediction for the mixed group of males and females by the RF algorithm using one hybrid model was feasible. The establishment of this model laid the foundation for thermal comfort-driven HVAC control methods and the improvement of indoor environments.