Machine Learning Predictive System Based on Transaxial Mid-femur Computed Tomography Images

Nowadays Machine Learning algorithms are commonly used in healthcare applications in order to help physicians in diagnosis or to find possible relations between measured biomedical parameters. In this thesis, starting from the AGES-Reykjavik Database, predictive analysis is done using both regression and classification ML algorithms. The Database is composed of 11 NTRA parameters extracted from Computed Tomography (CT) scans of mid-femur section of a 65-95 years old population, (4 related to muscle tissues, 4 to the fat, and 3 to the connective tissues) and by 25 measurements of which the most relevant are Body Mass Index (BMI), Cholesterol (SCHOL) and LEF biometric parameters (normal/fast gait speed, time up-to-go, and isometric leg strength). There are 3157 patients in AGES I and the same number in AGES II (same measurements on the same patients taken 5-6 after AGES I), so in total 6314. After an accurate study of the outliers and of the Nan values in the database, regression and classification are applied in order to predict at first BMI (parameter used to test the methodology) using tree-based regression algorithms like Decision Tree, Random Forest, Extra Trees, Ada Boosting and Gradient Boosting. Then the algorithms are extended also to SCHOL and LEF measurements. A proper Train-Test division of the dataset is done using k_fold cross-validation. Different selections of initial features are used combined with the different databases, with different k-fold divisions and using the 4 algorithms in order to obtain the best coefficient of determination (R2) as an evaluator of the quality of the regression’s prediction. Afterward, BMI and the other measurements are divided into 3 and 5 classes and the same methodology is used to classify them. For this classification analysis, the accuracy parameter (Jaccard Index - JI) is calculated to evaluate the quality of the results. The best results are obtained for the test parameter BMI. The max R2 for BMI is 0,8305 and it is obtained using as regressors the 11 NTRA parameter, AGESI+II, k_fold=16, and the Gradient Boosting Algorithm with n_estimators=200. For each R2 value obtained mixing all the possible combinations, the amplitude of the connective tissue and of the fat always cover more than 50% of all the feature importance. The classification gives a max JI=0,797 for 3 classes, while with 5 classes the max JI is 0,741. For what concern SCHOL, Normal Gait, and Time up-to-go, R2 and JI are not good enough to allow you to consider NTRA parameters as predictive features for them. On the other hand, the results of regression and classification for Leg Strength are satisfactory with a maximum R2 of 0,613 obtained using a particular selection of NTRA features, AGES-I, k_fold=16, and Gradient Boosting Algorithm. The best JI for Leg Strength is obtained with the division in 3 classes and is equal to 0,638. In conclusion, this study provides good results in terms of prediction, using both regression and classification for the BMI and the Leg Strength starting from a CT scan and adding eventually other physiological parameters. The methodology and the tree-based algorithms which are effective for BMI and Leg Strength can be eventually, in future works, extended also to other parameters present in the database.