Ton J. Cleophas

Machine Learning in Medicine - a Complete Overview

The current book is the first publication of a complete overview of machine learning methodologies for the medical and health sector. It was written as a training companion, and as a must-read, not only for physicians and students, but also for any one involved in the process and progress of health and health care. In eighty chapters eighty different machine learning methodologies are reviewed, in combination with data examples for self-assessment. Each chapter can be studied without the need to consult other chapters. The amount of data stored in the worlds databases doubles every 20 months, and clinicians, familiar with traditional statistical methods, are at a loss to analyze them. Traditional methods have, indeed, difficulty to identify outliers in large datasets, and to find patterns in big data and data with multiple exposure / outcome variables. In addition, analysis-rules for surveys and questionnaires, which are currently common methods of data collection, are, essentially, missing. Fortunately, the new discipline, machine learning, is able to cover all of these limitations. So far medical professionals have been rather reluctant to use machine learning. Also, in the field of diagnosis making, few doctors may want a computer checking them, are interested in collaboration with a computer or with computer engineers. Adequate health and health care will, however, soon be impossible without proper data supervision from modern machine learning methodologies like cluster models, neural networks, and other data mining methodologies. Each chapter starts with purposes and scientific questions. Then, step-by-step analyses, using data examples, are given. Finally, a paragraph with conclusion, and references to the corresponding sites of three introductory textbooks, previously written by the same authors, is given

Modern Meta-Analysis

A meta-analysis is a systematic review with pooled outcome data. The current chapter gives a summary of methods for the purpose. Four scientific rules and three pitfalls are reviewed. The main benefit of meta-analysis is that it provides a pooled outcome with increased precision. The main criticisms are that so far they were not good at predicting subsequent large trials, and at predicting serious adverse effects of medicines.

Validating Surrogate Endpoints

The International Conference of Harmonisation (ICH) Guideline E9, Statistics Principles for Clinical Trials, recommends that surrogate endpoints in clinical trials be validated using either (1) the sensitivity-specificity approach or (2) regression analysis. The problem with (1) is that an overall level of validity is hard to give, and with (2) that a significant correlation between the surrogate and true endpoint is not enough to indicate that the surrogate is a valid predictor.

Hierarchical Cluster Analysis for Unsupervised Data

Drug efficacy is multifactorial and with multiple variables regression modeling rapidly looses power and it is invalid if the correlations between the variables is strong. Hierarchical cluster analysis can handle hundreds of variables, and is unaffected by strong correlations.

Artificial Intelligence, Multilayer Perceptron Modeling

Back propagation (BP) artificial neural networks is a distribution-free method for data-analysis based on layers of artificial neurons that transduce imputed information. It has been recognized to have a number of advantages compared to traditional methods including the possibility to process imperfect data, and complex nonlinear data.

Regression Analysis in Medical Research

The current chapter reviews the general principles of the most popular regression models in a nonmathematical fashion. First, simple and multiple linear regressions are explained as methods for making predictions about outcome variables, otherwise called dependent variables, from exposure variables, otherwise called independent variables. Second, additional purposes of regression analyses are addressed, including 1. an exploratory purpose, 2. increasing precision, 3. adjusting confounding, 4. adjusting interaction. Particular attention has been given to common sense rationing and more intuitive explanations of the pretty complex statistical methodologies, rather than bloodless algebraic proofs of the methods.

Modern Bayesian Statistics in Clinical Research

Modern Bayesian statistics in clinical research , Modern Bayesian statistics in clinical research , کتابخانهu200cهای دانشگاه کردستان

Poisson Regression for Outcome Rates (50 Patients)

Poisson regression is different from linear en logistic regression, because it uses a log transformed dependent variable. For rates, defined as numbers of events per person per time unit, Poisson regression is very sensitive and probably better than standard regression methods.

Discrete Wavelet Analysis

Wavelets are oscillations, supposedly resulting from multiple smaller wavelets, and they are, traditionally, analyzed with polynomial, sine and cosine, and other functions. Ingrid Daubechies (1988) demonstrated that the repeated use of sharply spiked functions with multiple scales as basis functions for wavelet analysis provided better data-fit, and called it discrete wavelet analysis.

Multivariate Analysis of Time Series

Time series are encountered in every field of medicine. Traditional tests are unable to assess trends, seasonality, change points and the effects of multiple predictors like treatment modalities simultaneously. Autoregressive integrated moving average (ARIMA) is able to do all of that, but is, virtually, unused in medicine.