Signal Processing and Machine Learning for Biomedical Big Data
Ervin Sejdic and Tiago H. Falk, CRC Press, 2018, ISBN: 9781351061223, xviii+ 606 pages, $220
But for one flaw in the first printing of this text (to be detailed at the end of this review), this reviewer highly recommends this reference text. This 28-chapter, 89-author, large-format (8.5 × 11″), liberally illustrated (216 b/w figures), well-referenced and up-to-date text will be of value to those entering the field and likely to those interested in teaching various aspects of biomedical big data collection and analysis.
While defined elsewhere, chapter 24, titled “Trauma Outcome Prediction in the Era of Big Data: From Data Collection to Analytics,” nicely exemplifies the term “biomedical big data” by detailing the data generation of about 20 million data points per day per bed in a trauma center composed of 13 trauma resuscitation units, 9 operating rooms, 12 postanesthesia care units, and 60 bed intensive care units. Data collection includes (minimally) ECG, PPG, CO2, ABP, and ICP waveforms, and items such as interventions, events, radiological images, patient demographics, medical records, calculated indices, etc. Data outputs (obviously) include trends, warnings, real-time displays of waveforms, etc. The goal here (as in the remainder of the text) is to optimally monitor and intervene as possible in the treatment of patients.
This text comprises three sections: I “Introduction” (3 chapters), II “Signal Processing for Big Data” (9 chapters), and III “Application of Signal Processing and Machine Learning for Big Biomedical Data” (16 chapters.)
The editors, in chapter 1 of the introduction, give an overview of the text as to: 1) sources of big data (EHRs, imaging, genomics, wearables, biomedical instrumentation, etc.); 2) data qualities (volume, velocity, variety, veracity, validity, variability, visualization, and value) and the ability to analyze data in the face of noise and dropouts; with the 3) goals of saving money (an estimated $300 billion U.S. savings) and lives. Chapter 2 elaborates health care data sources (EHRs, clinical studies, outcome data, registry entries, genetic data, data sensors, and social media), drivers for change (Affordable care act, insurers, etc.), benefits (precision medicine, decreased ER use, etc.), and devices [monitors (home and HC facility), POC tests, etc.]. Chapter 3 then overviews the prior material as applied to analyses of MRI images, including machine learning from images, with the goal of ultimate clinical application to neurological and psychiatric disease diagnoses.
Nine chapters then discuss various signal processing methods for big data analysis. These range from straightforward time-frequency analyses of EEG data to “Targeted Learning” techniques to Bayesian State Estimations for big data. Dependent on your use of this text, this section should address your problem and its analysis, or can be used as an overview of methods for analysis of big data in a short course.
To this reviewer, the most useful aspect of this text is the collection of 16 chapters addressing “Application of Signal Processing and Machine Learning for Big Biomedical Data.” Chapters cover applications in: epilepsy, EEG interpretation, Image segmentation, retinal analysis, genomic analysis, sleep/insomnia studies, medical problem solving, fall risk assessment, trauma unit uses (see above, chapter 24), etc. Each chapter is well done, but an exemplar chapter (24) highlights succinctly one of the best reasons for a text such as this: While discussing the use of these techniques in data analysis, the datum is given that for an annual 200,000 ICU admissions in the U.S., there are only 40,000 discharges.
Excellent text, very well referenced, useful as a reference text or possibly as a course textbook for the mathematically gifted interested in all the methods and applications covered.
The one drawback mentioned above? The printed text that this reviewer was sent had more than 20 misprinted figures (some minor, some major) as a result of incorrect translation from author-supplied figures to the printed textbook. While some 20 of 216 may be minor, it can be irritating if the figure is one that you might need. The e-book version might be your best option for this text.