Developing an Algorithm to Identify Palliative Radiation Treatments: A Single-Institution 10-year Experience

Abstract

The definition of palliative radiotherapy (RT) has been changing over the past 20 years as patients live longer and radiotherapy for durable local control is delivered more frequently, even in the setting of metastatic disease. As such, the role of palliative RT continues to expand with changing and emerging treatment paradigms. The University of Pennsylvania established a dedicated palliative RT service in 2012 and has delivered several thousand courses of palliative RT over the last 11 years. In spite of a checkbox for “curative” or “palliative” intent, these courses are not always clearly defined in our treatment planning software. The development of a database of patients treated with palliative RT will allow us to study trends in approach and the algorithm may help other institutions to similarly study patients treated with palliative radiation. We sought to design a large, single-institution retrospective database of patients treated with palliative RT using an algorithmic approach to define and capture such patients based on clinical and electronic medical record data. All non-pediatric RT courses delivered at the University of Pennsylvania from 2008-2019 were eligible for review. Clinical details available for each course from the Varian Aria database (Varian Medical Systems, Palo Alto, CA) were course number, course name, provider name, treatment modality (proton vs. photon), treatment intent (provider specified as palliative vs. curative, and often left blank), associated ICD diagnosis code, prescribed dose, number of initial fractions, and number of cone down fractions. We pre-specified a target 95% sensitivity for identifying non-curative courses to include in the database. We defined palliative RT as any course delivered with non-curative intent including any ablative courses for oligometastatic disease, treatment for recurrent disease, or dose-escalated courses for durable local control. Investigators initially reviewed 214 courses delivered in January 2017 and based on chart review identified which courses were administered with curative versus palliative intent. This was repeated for courses delivered in March and April 2017. In total 634 courses were initially classified. We developed initial rules to reflect treatment parameters indicating an RT course was definitely curative (18 rules), definitely palliative (21), likely curative (12), or likely palliative (6). These rules were run on additional months as test data sets and adjusted to yield a higher sensitivity in identifying non-curative courses. Crude analyses were then iteratively run to determine which order and combination of rules led to best sensitivity in identifying palliative courses compared to gold standard of manual review of each treatment course. Our algorithm parses treatment courses in the following stepwise fashion: 1) Exclude courses delivered for benign conditions indicated by diagnosis code 2) Isolate all courses indicated by physician as “palliative” intent and add to database. Based on chart review “curative” was not a reliable indicator of curative intent as we have defined it. 3) From remaining courses isolate all courses with course names including “liver” and “SBRT” (stereotactic body radiotherapy) lung for manual review. The intent of these courses could not be reliably determined from available clinical details. 4) From remaining, exclude all courses with 1 or more “definitely curative” indicators and 0 “definitely palliative” indicators. 5) From remaining, isolate all courses with 1 or more “definitely palliative” and 0 or 1 “definitely curative” indicators and include in database. 6) From remaining, compare sum of “likely palliative” indicators and if greater than sum of “likely curative” indicators, include in database. With this basic approach we screened 23,782 courses and according to the stepwise process above found 1) 5,301 palliative courses, 2) 2,415 requiring manual review, 3) 13,467 excluded as definitely curative, 4) 2,195 definitely palliative, and 5) 404 likely palliative. In sum 7,900 palliative courses would be included in the database, with potentially more after manual review of 2,415 courses of liver-directed RT or lung SBRT. We were able to use a small number of relevant clinical details to reliably classify a majority of RT courses as curative or non-curative in intent. In the evolving setting of oligometastatic/oligorecurrent/oligoprogressive disease, identification of which RT courses are “curative” is highly variable between providers. Manual review of RT courses and rule generation results in >95% identification of palliative RT courses, but we have yet to show if machine learning techniques can improve on our manually generated algorithm.