Challenges in translating mortality risk to the point of care

Despite advances in medicine, prognostication remains inaccurate for many patients. Physicians tend to overestimate survival, even in advanced cancer and terminal illness groups.1–3 Over half of terminally ill patients express they do not want prolonging of life if their quality of life would decline.4 End-of-life interventions such as advanced care planning have shown improved adherence to patient’s wishes, improvement in satisfaction and reductions in stress, anxiety and depression,5 but clinicians remain reluctant to initiate end-of-life discussions with terminal patients if they are currently asymptomatic.6 Automated systems can complement clinician judgement to prompt earlier end-of-life discussions. To this end, predictive analytics is potentially impactful. Many different approaches have been used to estimate mortality risk using factors including severity of illness,7 healthcare utilisation8 or comorbidities.9 However, few works focus on palliative or end-of-life care (PEOLC), and even fewer have translated beyond model validation into prospective testing ultimately affecting clinical care. Instead, PEOLC remains reliant on clinical staff, despite their optimism, for initiation and prioritisation. The paper by Wegier and colleagues10 in this issue introduces a new 1-year mortality score—modified Hospitalised-patient One-year Mortality Risk (mHOMR)—designed for broad application at the time of admission. They incorporate mHOMR into two electronic health records (EHRs) to automatically identify patients who may benefit from palliative assessment. Of concern, there is evidence of patient distributional shift at the one site that showed improvement with the intervention. The authors conclude there was an increase in patients who receive palliative care consultations or goals-of-care discussions. However, the preintervention group appears much healthier, with a 3% in-hospital mortality, compared with the postintervention group (16%). Relatedly, a concomitant shift in patient mix to fewer frail patients is reported (68/100 to 43/97, p=0.001; Pearson’s χ2 test with Yates’ continuity correction). It is possible, therefore, …