Applying Machine Learning in Healthcare

A statistical method for analyzing a dataset in which there are one or more independent variables that determine an outcome – measured with a dichotomous variable (only two possible outcomes). In logistic regression, the dependent variable is binary or dichotomous, i.e. it only contains data coded as 1 (TRUE, success, pregnant, etc.) or 0 (FALSE, failure, non-pregnant, etc.).

A supervised machine learning algorithm that can be employed for both classification and regression purposes. SVMs are most commonly used in classification problems – the idea of finding a hyperplane that best divides a dataset into two classes, as shown in the image below.

An entire family of algorithms, all based on the idea of creating a tree of decisions about features that lead to a specific classification. Decision trees build classification or regression models in the form of a tree structure — breaking down a dataset into smaller and smaller subsets while developing an associated decision tree incrementally at the same time. The final result is a tree with decision nodes and leaf nodes. A decision node has two or more branches. Leaf nodes represents a classification or decision. The topmost decision node in a tree corresponds to the best predictor called a root node. Decision trees can handle both categorical and numerical data. There are also plenty of ways in which decision trees work for additional analysis like splitting, pruning, etc.

Deep learning (and neural networks in general) can take raw data as input and produce a class (or a vector of probabilities for many classes) as output. All neural network models consist of multiple layers of “neurons.” Each neuron in a layer receives the outputs of neurons in previous layers, combines these inputs, and uses a threshold to determine whether to output a value closer to 0 or closer to 1 for processing by the next layer. Deep-learning algorithms are unique in their ability to automatically generate features of interest in input data, as long as they are provided with a sufficient number of training examples (usually in the millions. although this is changing with new advances in the area of Transfer Learning). Deep learning is already used extensively in image, video, and audio understanding in healthcare, and it will eventually become more common in other classification problems in healthcare as larger sets of labeled training data become available for healthcare applications.

Associative memory systems compare incoming data with past data to create a pattern resemblance and to understand what the incoming data is all about. Comparisons can be based on any number of data attributes and have a lot of flexibility in that aspect. This is especially useful in healthcare where we typically have a large number of attributes. There are many measurements for each curation and each test generally has numerous data points.

Used for predicting the likelihood of categorical values such as diagnosis or special outcomes. Can also be used for complex applications such as probability estimation; e.g., Probabilistic Graph Models (PGM).

Deep learning systems using CNNs (Convolutional Neural Networks) have the capability to identify the hidden patterns inside images and videos. Deep learning algorithms have the capability to analyze images and what’s in them, without explicitly telling them what is crucial in it. Deep learning models need to be trained by being shown millions of images, each labelled with the desired variable marked. This training process is computationally intensive but once its trained, it can be used to easily identify the patterns.

HMM is a statistical Markov model in which the system being modeled is assumed to be a Markov process with unobserved; i.e. hidden states. These models are best known for their application in reinforcement learning and temporal pattern recognition such as speech, handwriting, gesture recognition, part-of-speech tagging, musical score following, partial discharges and bioinformatics.

LSTM models are similar to Convolutional Neural Networks described above. They can automatically learn what attributes of an audio stream are important for predicting what words they represent. Given sufficient data, which is readily available via online service providers such as Google and Baidu, it is possible to train LSTM models to accurately convert spoken language into text in almost any language. This technology has become the state-of-the-art for spoken language understanding applications and will likely play an increasing role in clinical transcription applications.

Unsupervised clustering algorithms such as K-means can logically group data based on their mutual distance and create centroids to create groups. This process is based on different parameters of data which can successfully divide the data into logical groups. This is a highly explainable algorithm which can be important in healthcare to explain how conclusions are reached.

Hierarchical clustering involves creating clusters that have a predetermined ordering from top to bottom. For example, all files and folders on the hard disk are organized in a hierarchy. There are two types of hierarchical clustering, Divisive and Agglomerative.