Deep Learning, a Not so Magical Problem Solver: A Case Study with Predicting the Complexity of Breast Cancer Cases

Area of Science:

• Oncology
• Medical Informatics
• Machine Learning in Healthcare

Background:

• Guideline-based clinical decision support systems (CDSSs) enhance cancer patient management within multidisciplinary tumour boards (MTBs).
• MTBs face challenges with overcrowding and limited time, necessitating prioritization of complex cases.
• Identifying complex breast cancer cases is crucial for optimizing MTB workflow and patient care.

Purpose of the Study:

• To develop a predictive model for identifying complex breast cancer cases (defined as non-compliant with CDSS recommendations).
• To optimize the workflow of multidisciplinary tumour boards (MTBs) by prioritizing complex cases.
• To evaluate the performance of machine learning algorithms in predicting non-compliant cases despite CDSS use.

Main Methods:

• Utilized Multi Layer Perceptron for classification tasks.
• Employed various sampling techniques to address data imbalance.
• Implemented cross-validation, hyperparameter tuning, and feature selection for model optimization.
• Defined complex cases as those non-compliant with the OncoDoc decision support system.

Main Results:

• Machine learning algorithms, including Multi Layer Perceptron, demonstrated insufficient performance in predicting complex cases.
• Despite extensive model optimization and data balancing techniques, the best achieved F1-score was 31.4%.
• The study highlights the difficulty in accurately predicting non-compliant cases within the current framework.

Conclusions:

• Current machine learning approaches are inadequate for accurately predicting complex breast cancer cases requiring prioritized MTB discussion.
• The poor performance suggests limitations in the definition of complexity or the predictive power of available features.
• Further research is needed to develop more effective methods for identifying and prioritizing complex cancer cases in MTBs.