Improving Breast Cancer Risk Prediction With Artificial Intelligence
Artificially intelligent (AI) systems are diagnosing cancer and other conditions with an accuracy that rivals that of trained pathologists. Research techniques using AI have produced new insights into the human genome, improved imaging techniques, and accelerated the discovery of new pharmaceutical drugs.
Much of this progress has been driven by advances in machine learning (ML) techniques, particularly computational analysis and statistical modelling, which have made the leap in recent years from the academic setting to everyday real-world applications.
Understanding risk factors
Comprehensive breast cancer risk prediction models generate absolute risk estimations to support clinical decision-making. They aim to classify women into clinically meaningful risk groups and enhance the identification and targeting of women at high risk while reducing interventions for those at low risk. Several mathematical and statistical models have been developed since the 1990s, based on large cohort datasets from different geographic regions, methodologies, and sets of meaningful risk factors (e.g., family history, genetic and epidemiological factors).
A major challenge in breast cancer risk prediction is to develop a model that incorporates all known and newly discovered risk factors while considering the interactions among them. In recent years, it has become increasingly clear that some exposures that affect breast cancer risk, particularly body mass index (BMI), nulliparity, and age at first live birth, change in magnitude and, in some cases, direction of effect over time.
ML techniques are perfectly suited to inferring and classifying such complex data interactions and continually optimising the learning process. ML offers an alternative approach to standard risk modelling that can address current limitations and improve the accuracy of breast cancer prediction tools. Indeed, advanced statistical analysis allows for a much more detailed exploration of the data and aims to improve the accuracy of the scoring algorithms over time as the amount of data used to train the underlying algorithms increases and the learning process is reinforced.
From black to white boxes
There are a number of social and ethical questions that must be raised when applying AI to real-world problems, especially in the fields of medicine and life sciences. Unlike human interpretation, it can be difficult to understand how or why an algorithm has made a certain decision or generated a particular result, known collectively as the black-box paradigm.
White-box models, on the other hand, have understandable behaviours, features, and relationships between influencing variables and output predictions. For example, simpler forms of ML, such as decision trees and Bayesian classifiers, that have certain amounts of traceability and transparency in their decision-making can provide the visibility needed for critical AI systems without sacrificing too much performance or accuracy.
Visibility is more difficult with complicated, more powerful algorithms, such as many-layered deep learning networks and stacked ensemble methods. Nevertheless, adoption of the white-box approach is clearly the logical way forward, especially for AI systems that predict or make decisions in a medical or healthcare setting. Moreover, the ability to correctly interpret model predictions engenders appropriate user trust, provides insight into how a model may be improved, and supports a better understanding of the process being modelled.
“Early diagnosis is the holy grail of breast cancer treatment.”
— MyBOOBRisk
