Harnessing Natural Language Processing and High-Dimensional Clinical Notes to Detect Goals-of-Care and Surrogate-Designation Conversations

Advance care planning, involving goals-of-care and surrogate-designation conversations, is crucial for patient-centered care. However, determining the optimal timing and participants for these conversations remains challenging. This study explored the frequency, timing, and predictors of documenting two advance care planning elements, goals-of-care and surrogate-designation conversations, in clinical notes for patients with advanced illness. In this retrospective observational study, we leveraged high-dimensional data and natural language processing (NLP) to analyze clinical notes and predict the presence or absence of advance care planning conversations. We included notes for patients treated at a Midwestern United States hospital who had advanced chronic conditions and eventually passed away. We manually labeled a gold-standard dataset (<i>n</i> = 913 notes) for the presence or absence of advance care planning conversations at the note level, achieving excellent inter-annotator agreement (90.5%). Training and testing four NLP models to detect goals-of-care and surrogate-designation conversations revealed that a transformer-based model (Bidirectional Encoder Representations from Transformers [BERT]) achieved the highest accuracy, with an F1 score of 93.6. We then deployed the BERT model to a high-dimensional corpus of 247,241 notes for 4,341 patients and detected goals-of-care and surrogate-designation conversations in the records of 85% and 60% of patients, respectively. Temporal analysis revealed that goals-of-care and surrogate-designation conversations were first documented at medians 28 and 8 days before death, respectively. Patient characteristics and referral to specialty palliative care emerged as significant factors associated with documenting these conversations. Our findings demonstrate the potential of NLP, particularly Transformer-based models like BERT, to accurately detect goals-of-care and surrogate-designation conversations in clinical narratives. This study identified significant temporal patterns, including late documentation, and patient characteristics associated with these conversations. It highlights the value of high-dimensional data in enhancing our understanding of advance care planning and offers insights for improving patient-centered care in clinical settings. Future research should explore the integration of these models into clinical workflows to facilitate timely and effective advance care planning discussions.