Abstract
Access to palliative care (PC) holds significance for hospital-admitted patients grappling with the symptoms of life limiting illnesses. Nonetheless, numerous such patients who could gain from PC fail to receive it promptly or even at all.
We can leverage the prior year's historical data extracted from electronic health records of hospitalized patients to train a machine learning (ML) model. This model's purpose would be to prognosticate the requirement for PC consultation using real-time data. The model, operating as a semi-supervised system, will be integrated into institutional data pipelines, and utilized by a downstream display application overseen by the PC team. In cases where the PC team deems it suitable, a team member will communicate with the respective care team of the patient. The ML model's training efficacy will be assessed using the area under the curve (AUC) metric, employing a 20% reserved validation set. The threshold for PC consultations will be grounded in historical data. To enhance the ML model's precision, the pivotal variables within the model will be pinpointed, and any sources of biases or errors in the model will be identified for meticulous refinement. The AUC values of successive ML models will be juxtaposed with cross-validation data.
Automatizing the referral procedure through electronic health record systems has the potential to usher in a more effective and streamlined approach to healthcare delivery.
Keywords
hospitalized patients - palliative care referrals - machine learning - predictive modeling - electronic health records
