To design a sensor prototype that emits multiple-wavelength lights and uses machine learning algorithms to measure hemoglobin noninvasively, and to assess the accuracy of its measurements. Methods: Eighty patients with ischemic stroke were enrolled consecutively. Venous blood was extracted to obtain the invasive hemoglobin concentration. Photoplethysmography (PPG) data were obtained simultaneously using a prototype finger clip device with three light-emitting diode (LED) sensors, and relevant features of the PPG signal were extracted. The emission wavelengths of the three LED sensors were 660 nm, 810 nm, and 1300 nm respectively. The features data sets of 40 patients were randomly selected for use in machine learning algorithm training, and the data sets of the other 40 patients were used for algorithm accuracy verification. Results: The hemoglobin concentration of patients ranged from 74 g/L to 177 g/L. The Pearson correlation coefficient between the hemoglobin concentration predicted by the non-invasive device and the results of invasive hemoglobin was 0.69, and the root mean square error was 13.19 g/L. The mean difference was (－3.2±12.95) g/L. Conclu? sion: The non-invasive device combined with multiple LED sensors and machine learning algorithm is feasible as a method for continuous hemoglobin level monitoring.