Machine Learning Classifier to Reproduce Lung Metastases Tumor Board Decisions

Abstract

Listen

Translate article

<h3>Purpose/Objective(s)</h3> Multidisciplinary tumor boards (MDTs) are an opportunity for specialists to meet and discuss the diagnosis and management of cancer patients. In our institution, patients with lung metastases are reviewed at a specialized tumor board comprising radiation oncologists, medical oncologists, surgeons, and radiologists. During these meetings, treatment recommendations for patients are made after a careful review of the patients' clinical history and relevant investigations. The objective of this study is to develop a machine learning classifier that reproduces MDT decision making and can be used as a decision support system to facilitate quality assurance and knowledge sharing in future MDTs. <h3>Materials/Methods</h3> Treatment decisions were manually gathered from patients discussed by the lung metastases MDT from 2016 to 2020 through clinical notes in the electronic medical records. Twenty-two variables were collected, including: age, gender, primary cancer histology, the number of lesions, the size of lesions, the presence of mediastinal/hilar lymphatic spread, the number of lobes/anatomical regions involved, disease free interval, ECOG status, and the presence of extra thoracic metastases. The target variable was binary: local therapy (SBRT or surgery) vs non-local therapy (systemic therapy, surveillance and palliative radiation). The dataset was split into a training, validation, and test set using a 60/20/20 split. Multiple machine learning models were evaluated (logistic regression, decision tree, support vector machine, naïve Bayes, random forest) with a combination of feature engineering, feature selection, and hyper parameter tuning. <h3>Results</h3> The dataset contains treatment decisions for 506 patients. The ages of patients ranged from 17 to 92 years old, with 215 female and 291 male patients. The most common primary cancers included colorectal (105), sarcoma (102), and head and neck (80). The dataset consisted of 235 local therapy, 111 systemic therapy, 152 surveillance, and 8 palliative radiation decisions. The best performing model was a random forest classifier, which classified local therapy vs non-local therapy decisions with an AUC of 86.6% and accuracy of 80.2%. The two most predictive features were the number of lesions and number of lobes/anatomical regions involved with metastases. <h3>Conclusion</h3> This study demonstrates that relatively simple variables taken from patient records can be used to create a model that generates recommendations concordant with an MDT. The next steps of the study include extracting and processing cross-sectional imaging to acquire exact tumor dimensions, density and location with respect to the pleura, fissures, bronchi, and vessels. This may improve the current classifier performance, as well as allow for more detailed predictions such as further classifying local treatment decisions into surgery and SBRT.