Abstract: Probabilistic Prognostic Estimates of Survival in Metastatic Cancer Patients

A separate "Palliative radiation dataset" was created using patients (899) enrolled from 2015-2016 in a prospective survey study conducted in our institution's Radiation Oncology department. The overall group of patients were seen for 471,005 daily encounters/visits, including outpatient and inpatient contact. For these visits, median followup was 12.7 months. Median overall survival was 22.4 months. Patients were hospitalized for 115,716 (24.6%) visits. There were 1,403,544 provider notes. The training set contains 10,239 patients with 380,080 visits, validation set of 1,785 patients and test set of 1,818 patients (15%) with 90,925 visits. Proposed System - PPES-Met: The model is composed of two core processing blocks: i. Semantic Word Embedding (SWE): We adopted a completely unsupervised hybrid method - an updated version of Intelligent Word Embedding (IWE) method(3) that combines semantic-dictionary mapping, neural embedding, and context-based windowing technique for creating dense vector representation of free-text clinical narratives. The method leverages the benefits of unsupervised learning along with expert-knowledge to tackle the major challenges of information extraction of informative information from clinical texts, while accounting for ambiguity of free text narrative style, lexical variations, arbitrary ordering of words, and frequent appearance of abbreviations and acronyms.