Hospice is a compassionate approach focusing on quality of life for terminally ill patients and their caregivers, with approximately 1.55 million Medicare beneficiaries enrolled in hospice care for at least one day during 2018 – 17% more than in 2014.
However, at least 14% of Medicare beneficiaries enrolled in hospice stayed for more than 180 days, and hospice stays beyond six months can result in substantial excess costs to healthcare organizations under value-based care arrangements.
David Klebonis, COO of Palm Beach Accountable Care Organization, has developed highly interpretable machine learning models that, because of the sensitivity of the clinical decision involved, cannot only accurately predict hospice overstays to drive appropriate hospice referrals, but also surface decision criteria that satisfy clinician scrutiny and promote adoption.
"Artificial intelligence and machine learning have the potential to use data to predict patients with a high probability of expiring within the next six months, so that physicians can enter into conversations with these patients and their families about the possibility of referral to hospice," he said.
Klebonis, who will address the topic this month at HIMSS22, said in Florida about 58% of Medicare decedents were in hospice at the time of death.
"If that number could be increased appropriately, it would be an improvement in quality of life for patients and their families while also reducing unnecessary healthcare costs," he said.
However, he added the most accurate AI/ML models in the world won't be adopted if physicians and other clinical end users don't understand or trust the predictions.
"It's critical to involve clinical stakeholders – that means the ones who choose whether to intervene based on the model's output – early in the ML development process," Klebonis said.
"They provide required context and feedback to help design a model such that it predicts outcomes that are clinically relevant and helpful and surface granular factors contributing to a prediction that can then give the clinicians data-driven insights they might not otherwise have had."
Another key element of building trust with clinical stakeholders is having rapid iteration cycles that allow you to build new models quickly, show the predictions and patient-level reports to clinicians for feedback and quickly incorporate that feedback in the next version of the model.
"To do this, it's important to have healthcare-specific tools that allow your team to minimize the amount of time spent on lower-level tasks like data cleaning and normalization, and instead focus on the outcomes, populations, and modeling features used to train a model and improve its accuracy," he said.https://www.healthcareitnews.com/news/explainable-ai-can-improve-hospice-care-reduce-costs