A Pilot on the Use of Unsupervised Learning and Probabilistic Modelling Towards Cancer Extent Prediction

Abstract

Cervical cancer is a global disease that attacks the female reproductive system and remains challenging to diagnose in the early stages. The rise in the application of machine learning models in various aspects of clinical medicine has seen research an increase in research applied toward the diagnosis of cervical cancer using a questionnaire style format. This work explores the application of unsupervised learning for an automated partitioning and learning of class labels, followed by subsequent learning of the designated labels using a supervised learning method. A contrast was conducted between two unsupervised learning methods, namely the agglomerative hierarchical clustering and k-means, where it was seen the k-means provided superior and more accurate clustering capabilities. The labels provided by the k-means were used to train the naïve Bayes (NB) classifier, which is capable of outputting probabilities associated with class predictions, therein allowing for the stage of cancer to be inferred. The results from the NB showed high accuracies in the region of 90+% across the various classifier metrics, therein showing good learning capability for learning from clustered data. The study provides evidence to show that a fully automated platform, which utilizes questionnaire style information, can predict whether a patient has cervical cancer alongside the likely cancer extent. Future work would now involve the use of a rule-based inference system that can automatically provide clinicians with a cancer extent where necessary and is also capable of handling further cancer stage subclasses information.

Received: 10 July 2022 | Revised: 24 October 2022 | Accepted: 25 October 2022

Conflicts of Interest

The author declares that he has no conflicts of interest to this work.