AI in Palliative Care and Mortality Prediction

Palliative care focuses on relief from symptoms and support for people living with serious illness; it is not limited to the final days of life. Some research models estimate mortality risk to help identify patients who may benefit from earlier conversations or specialist support. Published studies have developed and evaluated EHR-based or wearable models in specific populations, but results are tied to their datasets, outcomes, and health systems. NICE guidance on care in the last days of life notes uncertainty in recognizing when a person is dying and emphasizes clinical judgment and communication. A model’s mortality estimate should therefore not be presented as a definite timeline. It may prompt a team to review symptoms, care needs, and patient preferences, but it cannot determine what matters to a person or whether a referral is wanted. Validation should assess calibration, false positives and negatives, subgroup performance, and whether alerts lead to appropriate care. A high-risk score could trigger a compassionate conversation; it should not reduce access to treatment or be used as a stand-alone reason to limit care. Explain uncertainty, respect consent, and document the clinician’s reasoning. Evaluate outcomes such as timely conversations, symptom support, and unwanted burden, not only prediction accuracy. Track whether alerts widen access to specialist support or create unnecessary visits, and ask patients whether the conversation was helpful. Models should not force unwanted disclosure of prognosis; clinicians can tailor what is shared to the person’s preferences and decision-making needs.