Abstract
Dear Editor:
Like other medical specialties, palliative care (PC) generates tremendous amounts of data. The capture of these varied structured (e.g., quantitative measures such as the Palliative Performance Scale) and unstructured data (e.g., text outlining goals of care) presents immense opportunities to transform care for patients. For example, identifying “sentinel hospitalizations”—turning points in patients' disease trajectories that allow for meaningful end-of-life discussions—is a current gap. 1 We advocate embracing and accelerating adoption of state-of-the-art computational approaches, specifically machine learning (ML) and natural language processing (NLP), to uncloak these hidden opportunities to improve patients' quality of life and curtail expensive investigations and treatments that would have little benefit. In this viewpoint, we provide a brief overview of these approaches, highlight some of the ways in which these applications are currently being used, and discuss future directions.
ML and NLP
ML is an application within the field of artificial intelligence that leverages statistical techniques to allow computer systems to “learn” hidden patterns from data without being programmed. 2 NLP allows computers to process and analyze large amounts of unstructured language data, such as notes contained in electronic health records (EHRs), through rule-based or statistical-based approaches. 2 Although these applications have been gaining traction in other areas of medicine, they are slowly emerging in PC.
Current Examples in PC
At Stanford Hospital, Avati et al. developed a real-time model by applying a Deep Neural Network to EHR data to predict the mortality upon admission within 12 months to identify those patients who would benefit from a PC referral. 3 When the team manually reviewed patient charts in the model's test data, all were found to be appropriate for a consultation. 3 Udelsman leveraged a rule-based NLP model to improve the detection of patients with pneumoperitoneum and leptomeningeal metastases, serious sequelae of breast cancer. 4 Combining NLP with the existing process—use of administrative codes—resulted in a positive predictive value of 100%. 4
Future Directions
Although these are important advances, we need to build on their momentum. We must take advantage of a variety of data sources, including outpatient primary care, home visits, and social services, and apply models that incorporate both ML and NLP. Doing so will improve the performance of predictive algorithms, move us closer to enhanced prognostication of disease, and help clinicians have more informed conversations with patients and families.
Similarly, recognizing that PC is whole-person care delivered by clinical and nonclinical personnel working across the continuum of care, we need to construct user friendly workflows that tie predictive models to evidence-based interventions, translating high-value opportunities into better quality of life.
Finally, we must keep equity at the forefront of our work. We have to ensure that structurally vulnerable and disenfranchised patients are well represented in the data used to train algorithms and that they have access to services when appropriately referred. In this way, ours will be a system in which all patients receive high-quality palliative and end-of-life care when and where they need it.