Interpretation of Machine Learning Models for Medical Diagnosis

Abstract

Machine learning has been dramatically advanced over several decades, from theory context to a general business and technology implementation. Especially in healthcare research, it is obvious to perceive the scrutinizing implementation of machine learning to warranty the rewarded benefits in early disease detection and service recommendation. Many practitioners and researchers have eventually recognized no absolute winner approach to all kinds of data. Even when implicit, the learning algorithms rely on learning parameters, hyperparameters tuning to find the best values for these coefficients that optimize a particular evaluation metric. Consequently, machine learning is complicated and should not rely on one single model since the correct diagnosis can be controversial in a particular circumstance. Hence, an effective workflow should effortlessly incorporate a diversity of learning models and select the best candidate for a particular input data. In addressing the mentioned problem, the authors present processes that interpret the most appropriate learning models for each of the different clinical datasets as the foundation of developing and recommending diagnostic procedures. The whole process works as (i) automatic hyperparameters tuning for picking the most appropriate learning approach, and (ii) mobile application is developed to support clinical practices. A high F1-measurement has been achieved up to 1.0. Numerous experiments have been investigated on eight real-world datasets, applying several machine learning models, including a non-parameter approach, parameter model, bagging, and boosting techniques.

A Dropout Optimization Algorithm to Prevent Overfitting in Machine Learning

“…Search and predict the algorithm related to it and can improve the accuracy of model parameters through a certain program. The algorithm classifies the given data with the expected output as the constraint [9][10]. In reality, there are many uncertain factors, such as random interference, such as light intensity, temperature and other external environmental factors, and the above parameters themselves will produce errors that affect the final results.…”

IM2P-Medical: Towards Individual Management Privacy Preferences for the Medical Web Apps

“…Besides, Hoang et al [8] provide a mechanism to handle conflicts between the privacy policy and privacy preferences where it depends on prioritizing patient treatment or reducing the risk of personal information leakage. In addition to the above studies, the systems that build smart contract models for medical facilities using Blockchain technology also take care of users' privacy preferences, for example, Nghia et al [9], [10], [11]. In these studies, the patient role was given full discretion in sharing their data with stakeholders.…”

Classification-Segmentation Pipeline for MRI via Transfer Learning and Residual Networks

“…Identifying and analyzing diseases is increasingly difficult because they are ever more sophisticated. Fortunately, artificial intelligence has revolutionized clinical practice in many areas such as cancer diagnosis with medical imaging [4], automatic classification diseases based on descriptions [5], [6], and maximizing hospital efficiency [7]. Among many approaches, deep learning has been proven superior in a wide range of clinical data and practice scenarios.…”