Machine learning has emerged as a powerful tool in fields as diverse as cybersecurity and chess programming in order to automate redundant tasks or discern patterns quickly. The utility of machine learning tech is that it gets better over time with very little human input. As a result, humans shift their energy and attention to more pressing problems.
Interestingly it took a while for machine learning to appear in medicine. Problems like scaling a medical practice, or reducing errors in discerning anomalies in medical images are excellent use cases for machine learning, but hardware limitations and the regulatory landscape limited this technology’s practical applications.
That has all changed as the past decade has seen an explosion of machine learning deployment into healthcare. The keystone elements behind all of this disruption are the algorithms which power the machine learning models which provide practical solutions to healthcare professionals. It’s worth exploring what some of these algorithms look like with a few examples.
Natural Language Processing
A very large portion of clinical information is stored in text form. Healthcare relies on narrative text in the form of clinical lab reports, physical exam notes and discharge summaries. Oftentimes, the text in these documents is unstructured, which makes it impossible for computers to interpret. Natural Language Processing, which refers to the automatic manipulation of natural language, like speech and text, by software, addresses this issue in a couple of ways. One technique is to identify clusters of disease-relevant keywords in clinical notes and to compare them against a historical database. Another technique is the bag-of-words-model which is used to simplify text as string representations.
Natural language processing is used to predict patient outcomes, augment hospital triage systems, and generate diagnostic models that detect the onset of disease.
Convolutional Neural Network
Accurately classifying medical images is crucial for determining correct treatment for many medical illnesses. Analysis X-ray, for instance, is considered the best approach to diagnose pneumonia, but classifying pneumonia from chest X-rays is time and resource intensive.
Convolutional Neural Networks, or CNNs were developed to interpret high-dimensional data, such as images. According to Towards Data Science “A CNN algorithm learns by using multiple layers of “nodes” and numerous iterations of “analyzing” and “weighting” the structures, patterns, etc., it identifies in the images. A structure/pattern that helps it correctly identify or classify the image is “rewarded” with a higher “weight” while one that hurts or is useless is “punished” with a lower “weight.”
CNNs have already been shown to classify anomalies in medical images with great speed and accuracy, and they’re likely to get better. Widespread adoption of CNN technology would allow radiologists to more quickly detect anomalies and better serve their patients.
Naïve Bayes is a body of probability related algorithms that have proven powerful for categorizing text. These algorithms predict the probability of a set of categories for each piece of data, with the end result being the most likely category being assigned to the data.
Naive Bayes is relatively simple to implement and highly efficient. lt has been successfully applied to many medical problems, such as classification of medical reports and journal articles.