Machine learning, the name given to a vast class of methods that aim to enable machines to learn from data, is becoming more and more popular in many diverse fields, such as finance, the physical sciences, and medicine. It is practitioners in this latter field that Cleophas and Zwinderman’s book seeks to help. The book is a cut-down version of the much larger and more detailed volume Machine learning in medicine [1], and the general premise of the “cookbook” edition is simple: a set of “recipes” are provided that use machine learning to help solve a specific set of problems within medicine. So, for example, chapter 4 shows how regression can be used to help understand and extrapolate from patient sleep patterns. Unlike traditional textbooks, this book seeks to provide brief recipes for practitioners to use without knowledge of machine learning. But does such a book work?

The book contains 20 recipes, split across three categories: cluster models, linear models, and rules models. Each chapter (recipe) follows the same structure: it begins with a brief description of the general purpose of the chapter and is then followed by a description of the scientific problem. The bulk of each chapter is taken up with the core recipe and ends with a conclusion and further reading notes. Each recipe uses medical-specific data and a specific scientific question to illustrate the machine learning method discussed. All of the data and examples used in this book are available online through Springer’s extra materials portal [2]. This means that readers can follow through on all of the examples used in the book, as long as they have access to the SPSS [3] environment.

Although it is commendable that the authors have provided all data and detailed SPSS instructions, here lies one of the key problems with the book. It is heavily reliant on SPSS, to such a point that the title would be better rendered as Machine learning for medicine: an SPSS cookbook. The reader is told how to perform tasks such as regression, decision tree analysis, and clustering within SPSS, but is left in the dark as to the actual underlying workings of these techniques. They are presented as black boxes, with inputs on one side and outputs on the other; how the outputs are derived from the inputs is left entirely to the reader to find out. Now it might be unfair to criticize the book on this fact since, as previously mentioned, it is deliberately meant to be brief and lacking in detail. But after reading the book, I was left feeling that even a tiny bit more detail would help make it more widely applicable.

However, the heavy reliance on SPSS is not the main problem. My biggest concern is that each chapter is hyper-focused on the example given. This means that the recipes given are not easily generalized to new problems. By way of an example, consider the recipe given in chapter 5 where generalized linear models are used for outcome prediction with paired data. The step-by-step recipe details how the reader can use the example data to predict hours of sleep in groups and future patients. But what happens if we want to employ generalized linear models on a different dataset? There is scant information on the workings of the generalized models themselves, and the recipe given is engineered to the example data so it is not clear how to utilize the method for new data. Obviously, if readers have knowledge of how such models work, then they won’t have difficulty in applying the method to new data; if that is the case, why would they buy this book?

So does this book work? My feeling is that without the much larger parent volume [1], the effectiveness of this book is questionable. As a quick reference for those with access to the larger volume, I’m sure that this cookbook could prove useful. However, without a much larger text to fill in the gaps, this book will be of little use to readers who have no SPSS or some statistical and machine learning knowledge.

More reviews about this item: Amazon