Expert Consensus and Artificial Intelligence in Medical Decision Making in Patients with Malignant Brain Tumors

Nearly 23,000 adults are diagnosed with primary central nervous system (CNS) malignancy yearly. An additional 200,000 adults are diagnosed with brain metastasis. There are significant variations in CNS tumor treatment. However, due to significant heterogeneity in patient baseline factors, identifying unwarranted variation is challenging. Ghogawala et al have previously demonstrated that, among patients undergoing surgical treatment of cervical myelopathy and lumbar degenerative spinal disease, an expert panel consisting of surgeon experts can identify variations in proposed surgical procedure and demonstrated superior patient outcomes when the surgery performed matched the procedure recommended by expert consensus. Expert panel surveys have not previously been used to identify variations in care among patients with CNS malignancy. The primary aim is to determine whether patient outcomes are superior when treatment aligns with recommendations made by a clinical expert neurosurgical panel. The study also seek to identify patient factors that predispose to variability in care. Our long-term aim is to determine whether predictive artificial learning algorithms can achieve the same outcomes, or better, as clinical expert panels, but with greater efficiency and greater capacity to be available for more patients. The investigators hypothesize that: * When a team of 10 medical experts has greater than 80% consensus regarding optimal treatment and when the doctor and patient select that specific treatment, the outcome is superior than when a patient and doctor select an alternative procedure. * When a team of 10 medical experts has greater than 80% consensus regarding optimal treatment, the structured data used by the experts can be processed and trained by computing algorithms to predict the pattern recognized by the experts - i.e. - the computer can predict how an expert panel would vote. Procedures include the following: 1. Chart review portion of study: Patients will be identified from case logs of the principal investigators from July 2017 through July 2023. Data will be collected retrospectively and will include age, non-identifier demographics, diagnosis details, operative/treatment characteristics, post-treatment characteristics, and follow-up characteristics. Images reviewed will include pre and post-treatment MRIs obtained as part of routine care. Data will be abstracted from the medical record (Epic/Soarian and PACS) and recorded in an excel database. 2. Survey portion of study: De-identified structured radiographic data and a brief clinical vignette without patient identifiers will be uploaded to Acesis Healthcare Process Optimization Platform (http://www.acesis.com/our-platform). A survey will be generated by Acesis and emailed to the subject experts/participants. This portion is prospective. 3. Cohort definitions: 1. Patients will be assigned to either "expert-treatment consensus" or "no expert-treatment consensus" arms based on whether greater than 80% consensus is achieved 2. Patients will be assigned to either "Expert consensus-aligned" or "Expert consensus - unaligned" arms based on whether expert survey results match actual treatment given. 4. Data will then be analyzed using appropriate packages with SAS statistical analysis software. Survival analysis will be performed to determine whether consensus predicts improved progression free survival (PFS). 5. The structured and de-identified radiographic images used by the experts in surveys will be used for training and development of an AI algorithm. The aim of this portion of the study is to determine whether standardized and structured imaging can be used to train an algorithm to predict whether expert consensus is achieved and the recommended treatment.