Intelligible Models for HealthCare

High accuracy by machine learning models imply its ability to closely mimic data generating process but they may possess the property of low interpretability by humans (called Intelligibility).

Complex models do not explain their prediction well, which can act as a barrier to their adoption. For mission critical applications where interpretability helps in understanding the role of each feature contributing to the final outcome, complexity of models may hinder such causal effects.

A simple rule based model might not be as accurate as a neural network to learn that asthma patients or pregnant womens are less prone to death by pneumonia, but it’s easy to remove rules producing such generalisation and are hence editable.

Generalised-Additive-Modelling(GA2M) technique is both competitive and intelligible compared to random forest and logit boost algorithms by comparing their AUC score on two different real-world problems from the healthcare domain.

GA2M technique considers both contribution of single and pairwise interaction of features while modelling data. Top K pairwise interaction terms are chosen by ranking all possible pairs.

Pneumonia risk: Modelling dataset of predicting risk of pneumonia, GA2M suggests some other features such as asthma which negatively affect the learning…