Abstract
Abstract only
e24085
Background: Breakthrough cancer pain (BTcP), a transitory flare of pain that occurs on a background of relatively well-controlled baseline pain, is a challenging clinical problem in managing cancer pain. We hypothesized that the BTcP could be predictable according to the patients’ previous observed patterns. In this study, we report on the development of a deep learning model that predicts hourly individual-level breakthrough pain for patients with cancer. Methods: We defined the BTcP as the pain with numerical rating scale (NRS) score 4 or above and developed models predicting the onset time of BTcP with the temporal resolution of 1 hour. The datasets which have more than 20 records of NRS score during hospitalization were included in our study. All the pain records were obtained from patients hospitalized on the wards of hematology-oncology in Samsung Medical Center between July 2016 to February 2020. The model used the time windows of 3 days to predict NRS scores over the next 24 hours. To capture irregular pain patterns, we created the sequence of average pain patterns over 24 hours from the previous 3 days and used it for normalization. We trained a Bi-directional long-short term memory (LSTM) based deep learning model. The model was validated using the holdout method with 20% of the datasets. Its performance was assessed with the area under the receiver operating characteristic curve (AUROC) and the area under the precision-recall curve (AURPC). Results: We included pain log data containing 2,905 admissions from 2,176 patients with solid cancer and 1,755 admissions from 1,082 patients with hematologic cancer in the analysis. The median age was 57 (interquartile range (IQR), 47-64), the most frequent type of cancer was lung cancer (18.0%), and most patients had stage 4 (60.7%). Among the 103,948 hours from patients in whole datasets, 1,091 (4.7%) hours were labeled as the period of BTcP. The patients have the records of NRS score with a median of 3 (IQR, 2.0-4.5) and BTcP with a median of 1.1 (IQR, 0.5-2.0) per day. We allocated approximately 20% of patients (653 patients with 932 admissions) to the holdout test dataset. Our model showed the AUROC 0.719 and AUPRC 0.680 for predicting the BTcP in the test dataset. Conclusions: Our study showed that cancer pain could be predictive by using a deep learning model. Though our exploratory study has a limitation of generalizability, future warranted subgroup analysis and verification research could make our model more applicable in a real-world setting.