Cancer predictive model derived from bioimpedance measurements using machine learning methods

Summary: Objective: This work is aimed to develop a machine learning predictions of health status derived from bioimpedance measurements of adult healthy and cancer individuals. Methods: We conducted a pilot random study containing 2881 female (1220) and male (1661) patients ranging in age between 19 to 96 years old are studied. Among of them, 33 are diagnosed with cancer disease, the rest are healthy. After balancing the initial data, the data of interest contains 1.460 individuals ranging in age between 19 and 93 years old (734 female and 726 male), with 704 diagnosed with cancer and 756 healthy, respectively. The bioimpedance parameters are obtained by measuring standard tetrapolar whole-body configuration. The bioimpedance analyser (BioScan98®) is used, collecting fundamental bioelectrical and other parameters of interest. A classification model are performed, followed by a prediction of phase angle and body mass index. Results: The classification model reveal two robust parameters for predicting the health status, namely the impedance, the total body water and the phase angle with a 97%, 34% and 30 % of significance (respectively), with an area under the receiver operating characteristic curve of AUC = 1.00. The phase angle predictions agrees with previous reports of other type of pathologies, where higher phase angle values is ascribed to better health status and male have larger values than female. Recommendations regarding the capacitive reactance as a robust parameter to inferring health status is discussed. The cubic support vector machine model shows great accuracy predicting the nutritional status based on body mass index of both healthy and cancer patients. Conclusion: The classification, phase angle and body mass index predictive models developed in this work are of the great importance to assist the diagnosis, differentiating between healthy and cancer individual with great accuracy. Despite the moderate lack of body mass index association with cancer, this model can be used for prompt diagnosis.