Abstract
INTRODUCTION: Increased sophistication in machine-learning algorithms and artificial intelligence have begun to unveil patterns that would be otherwise unappreciated in clinical medicine. Here we applied one such algorithm, Iterative Factorial Analysis of Mixed Data (IFAMD), to better understanding combinations of clinical variables that influence clinical survival of brain metastasis (BM) patients treated with stereotactic radiosurgery (SRS). METHODS: A dataset of 6,326 BM patients was collated from four SRS centers (University of California, San Diego, Katsuta Hospital Mito GammaHouse, Tsukiji Neurological Clinic, and Melanoma Institute of Australia). IFAMD was applied to the analysis of the following clinical variables: age, Karnofsky Performance Status (KPS), cumulative intracranial tumor volume (CITV), total number of metastases, histology (breast, gastrointestinal (GI) cancer, renal cell carcinoma (RCC), melanoma, and lung cancer), systemic disease control, and survival in months. RESULTS: Our machine learning algorithm defined three groups of patients who exhibited differential survival. The group who is most likely to die within 3 months of SRS included patients with lower KPS, poor systemic disease control, higher CITV, higher number of metastasis, and who carried a diagnosis of GI cancer. Patients who are most likely to survive beyond twelve months of SRS fall into two distinct categories. The first consisted of subsets of lung and breast cancer patients with higher KPS, controlled systemic disease, and lower CITV. The second consisted of young breast cancer patients with systemic disease control, independent of KPS, CITV, and the number of metastases. CONCLUSION: Clinical survival after SRS for BM is defined by combinations of known prognostic factors. A prognostic factor critical for survival prognosis in one sub-population of BM patients may bear little relevance in another patient sub-population.
Building prognostic models for breast cancer patients using clinical variables and hundreds of gene expression signatures