Validated Machine Learning Models Show Significant Success

by Literature Review •

What are the capabilities and limitations of machine learning for healthcare datasets, particularly for otolaryngologists and head and neck surgeons?

BOTTOM LINE: Investigators have realized significant success in validated models using machine learning, but challenges remain in their implementation.

BACKGROUND: Machine learning is an artificial intelligence subset that is concerned with making predictions in novel situations using data from previous observations and is widely used in medicine to enhance cancer diagnosis and prognosis predictions by integrating clinical and genomic data, and adaptive clinical trial designs. Its application in medicine is relatively new, and there may be significant unfamiliarity in its use.

STUDY DESIGN: Literature review.

SETTING: Sunnybrook Health Sciences Center, Toronto, Ontario, Canada.

SYNOPSIS: Machine learning algorithms generally fall into four main categories: category prediction, structure prediction, dimension reduction, and value prediction. These approaches can be further categorized as supervised (known starting and outcome points) and unsupervised (unknown outcome/output) learning. An algorithm class gaining intense interest in medicine is deep learning, based on multilayered neural network models that dissect and distill complex information to a perceptible output. These techniques have been used in otolaryngology–head and neck surgery clinical problems, including predicting recovery/nonrecovery of adults with sudden sensorineural hearing loss (overall accuracy 77.6% [neural networks] versus 67.5% [logistic regression]); detecting differences between normal and malignant head and neck tissues using hyperspectral imaging (97% sensitivity, 96% specificity, 96% accuracy); and recognizing the differences between normal subjects and T1a glottic carcinoma vocal cord wave morphologies on video stroboscopy (100% sensitivity, 100% specificity). There is a certain unease, however, in relying on AI systems in which we are unable to pinpoint what rules and processes were used to generate output, and there are concerns about the potential for machine learning implementations to propagate error from imperfect measures.

CITATION: Crowson MG, Ranisau J, Eska-nder A, et al. A contemporary review of machine learning in otolaryngology–head and neck surgery. Laryngoscope. 2020;130:45-51.