MACHINE LEARNING APPLIED TO CLINICAL NUTRITION: CLINICAL DECISION SUPPORT AND NEW PATTERNS RECOGNITION

Introduction. The healthcare sector is failing to utilize routinely produced clinical data to refine the care experience and to augment knowledge created in clinical study. The big data culture and the closely connected field of machine learning constitute the latest and the best opportunity yet to put to good use data created as a by-product of clinical care. Aim of this thesis was to test the capabilities of machine learning algorithms applied to real-world clinical nutritional data to assist clinicians in their decision making process. Machine learning was used in two predictive contexts: 1) prediction of routinely collected parameters for patients non-eligible for the reference method, and 2) prediction of failure to meet clinical targets set out for the patient. Material and methods. A large nutritional dataset collected at the International Center of Nutritional Status (University of Milan, Milan, Italy) was used for the analysis. The dataset include 15780 patients and multi-domains predictors providing informations on age, sex, education, occupation, marital status, family status, menstruation, pregnancies, diet status, diet history, physical activity, smoking, pharmacological treatments, clinical signs, weight history, physical exam, blood pressure, anthropometry, bioimpedance analysis, ultrasound (abdomen fat thicknesses), indirect calorimetry, laboratory exams, anxiety, depression, binge eating, emotion regulation, eating disorders, and adherence to a Mediterranean diet. Machine learning algorithms were applied in order to predict the following outcomes: resting energy expenditure by indirect calorimetry, total body water by bio-impedance analysis, weight loss failure, failure to improve basal glycemia, failure to improve total cholesterolemia, failure to improve triglyceridemia. To evaluate accuracy and discrimination ability of machine learning and statistical algorithms, a series of cross-validation experiments were conducted for all outcomes, and the most accurate algorithm for each outcome was selected as the best for that outcome. Accuracy was defined with the root-mean-square error for continuous outcomes and the correct classification fraction for categorical outcomes. Results. Machine learning algorithms outperformed statistical algorithms for all outcomes. The best performing models were tree-based models, in particular bagged decision trees performed best for continuous outcomes, while random forests performed best for categorical outcomes (with the exception of the triglyceridemia outcome which saw a boosted tree algorithm as the best performer). In the prediction of resting energy expenditure and total body water, accuracy was high and the mean errors were deemed small in the context of clinical practice [mean (95% confidence interval) root-mean-square error 27.6 (20.9, 34.3) kcal/day and 0.842 (0.768, 0.916) l respectively]. In the prediction of weight loss failure, failure to improve basal glycemia, failure to improve total cholesterolemia, and failure to improve triglyceridemia, the mean correct classification fraction ranged between .616 and .735, but even the best algorithms showed good sensitivity but poor specificity (mean area under the ROC curve ranged between .652 and .687). For categorical outcomes unbalanced toward the event, machine learning models were the only one able to improve the accuracy of a naive classifier that assumes that all patients will experience the event, although only in weight loss failure model outcome accuracy was consistently above the naive classifier. Discussion. Our results highlight the ability of machine learning algorithms to provide a high-accuracy alternative to reference techniques for non-eligible patients. The big-data culture paired with machine learning algorithms seem able to overcome limitations imposed from using externally-developed equations, providing highly accurate predictions. In the setting of identifying non-responders, machine learning algorithms did not provide highly discriminant predictions, but were the only ones able to provide a better prediction of random guessing or the historical rate of event. In this more ambitious task, machine learning algorithm results need to be critically interpreted by the clinician, whose reasoning is necessarily different but can incorporate the suggestions provided from these algorithms.