Natural Language Processing of Symptoms Preceding Diagnosis and Palliative Radiotherapy for Bone Metastases

Abstract

<h3>Purpose/Objective(s)</h3> In patients with bone metastases (BM), the goal of palliative radiotherapy (RT) is to ameliorate pain and other symptoms. We sought to apply natural language processing (NLP) techniques to routine clinical documentation to characterize symptoms associated with diagnosis of BM and initiation of palliative RT. We hypothesized that there may be disparities in symptom presentation for patients with BM receiving palliative RT. <h3>Materials/Methods</h3> We used International Classification of Diseases (ICD) codes along with Current Procedural Terminology (CPT) codes to create a de-identified, single tertiary-care institution cohort of patients with BM who received palliative external beam RT or stereotactic body RT between 2013 and 2021. Clinical data, clinical notes, and meta-data on clinical notes for the patient cohort were computationally extracted from an institutional electronic health record repository. We applied a previously validated clinical Text Analysis and Knowledge Extraction System (cTAKES) pipeline to extract Common Terminology Criteria for Adverse Events (CTCAE) encoded symptoms from all notes in the 30 days preceding BM diagnosis and in the 30 days preceding each course of palliative RT. Chi-squared tests were used to compare symptom distributions across groups. <h3>Results</h3> There were 1,021 patients who received 1,551 courses of palliative RT for BM. Median age at BM diagnosis was 63 years (IQR: 54-71). Median time from BM diagnosis to first course of palliative RT for BM was 2.3 months (IQR: 0.7-14.9). Most patients were male (57.5%), English-speaking (86.2%), and white (59.2%). The most common symptoms prior to BM diagnosis were pain (75.6%), nausea (44.9%), fatigue (41.6%), and palpation of a mass (38.4%). Across RT courses, the most common symptoms prior to palliative RT were pain (90.4%), nausea (67.9%), fatigue (67.6%), and constipation (54.2%). On subgroup analysis of symptoms prior to RT, there were increased NLP-derived reports of pain in clinical notes among patients who were female (94.9% vs male: 87.0%, p < 0.001), non-white (93.3% vs white: 88.8%, p = 0.007), and non-English speaking (97.0% vs English: 89.5%, p = 0.002). Compared to their respective counterparts, female (33.2% vs male: 24.6%, p < 0.001) and non-English speaking patients (36.5% vs English: 27.2%, p = 0.012) were more likely to have experienced pathologic fractures prior to RT. <h3>Conclusion</h3> Patients who were women, non-white, or non-English speaking were documented to have higher rates of pain and/or pathologic fracture prior to palliative RT. These findings may reflect disparities in patient presentation, clinical documentation, or decision-making for which patients are offered palliative RT sooner in the disease trajectory. Future research coupling NLP techniques with machine learning may be useful for predicting which patients would benefit most from earlier initiation of palliative RT.