OBJECTIVES: Discharge summaries of hospitalized patients often contain unstructured information that is difficult to represent and analyze. In this study, we propose a machine learning solution for the automatic reading and labeling of medical instructions given to the patient at the time of hospital discharge. Elderly aged 90 years or older were selected as a case study, given their special need for medical assistance.

METHODS: The discharge summaries of 642 patients in 2021 were analyzed, containing 294K sentences. The database was submitted to a labeling process by a healthcare professional, classifying them into seven medical action items: appointment, lab, home care, medication, rehabilitation, recommendation, and transfer. 137K sentences (335 patients) were used in the models training and validation process. The model’s architecture was built from a pre-trained deep learning algorithm for natural language processing, connecting its outcome to the input of a classification algorithm for each label.

RESULTS: In the test dataset (157K sentences, 335 patients), the labels, highly unbalanced (11.6% of positive instances), were distributed in: medication (7.35%), recommendation (2.06%), appointment (1.34%), home care (0.72%), lab (0.08%), rehabilitation (0.04%) and transfer (0.001%). Given disproportionate classes problem, sensitivity and area under the ROC curve were the main accurate metrics. The highest sensitivity was observed in transfer at 100% (AUC 82%), followed by lab at 95% (AUC 84%) due to the lower number of instances. However, the wide vocabulary variation showed a lower sensitivity in recommendation at 83% (AUC 79%), although the lowest sensitivity was observed in home care at 72% (AUC 77%). The models showed an average sensitivity of 89% and AUC of 81%.

CONCLUSIONS: The case study led to the creation of a labeling tool for actionable items in discharge summaries, which supported a team of health professionals to follow the care trajectory of the elderly patient, providing better assistance to their health.