Electronic health record-based algorithms are being used to identify patients with high palliative care needs and trigger earlier action. Across seven randomised trials involving 125,666 patients, these systems were linked to higher rates of palliative care consultation and do-not-resuscitate documentation. The trials spanned inpatient wards, intensive care units, an emergency department and a community clinic, using different combinations of clinical and administrative data to automate identification and referral. Effects on hospice use and in-hospital mortality were marginal, while no significant effects were found for ICU admission, length of stay or family-reported psychological outcomes. The clearest impact was on consultation and advance care planning documentation.
Must Read: AI Closes Data Gaps for Rare ICU Conditions
Consultation Rates Rose Across Clinical Settings
Five trials with 26,701 patients contributed to the pooled analysis of palliative care consultation. Automated EHR algorithms produced significantly higher consultation rates than usual care, with a pooled risk ratio of 2.67. Subgroup analysis also showed significant effects in both cancer and noncancer populations. In cancer, the risk ratio was 5.31. In noncancer populations, the risk ratio was 2.19.
The included interventions used different operational models. Some systems generated default palliative care orders for hospitalised patients who met age and diagnosis criteria. Others used high-risk phenotypes, length of ICU stay, NEST thresholds, early warning alerts or control tower screening of high-need inpatients. In one cancer-focused setting, weekly EHR identification was followed by chart review and palliative care initiation within 48 hours after physician approval. In another emergency department programme, clinical decision support banners were combined with communication training, audit and feedback.
The common feature across these interventions was the use of routinely collected EHR data to identify patients who might benefit from palliative care earlier in the course of hospital care. Consultation was defined through documented specialist involvement such as signatures, meeting records or consultation notes. Despite substantial heterogeneity across the pooled consultation analysis, the direction of effect remained favourable in both major subgroups. Six of the seven included trials were rated as low risk of bias overall, while one had some concerns. Certainty of evidence for consultation outcomes was rated high.
Documentation Improved While Other Outcomes Remained Mixed
Two randomised trials with 15,898 patients examined do-not-resuscitate documentation. The pooled analysis showed significantly higher documentation in the algorithm group, with a risk ratio of 1.22. This result aligned with the broader pattern of increased palliative care activity after automated identification and referral.
Hospice use was assessed in 2 trials involving 99,101 patients. The pooled result showed no significant difference between groups, with a risk ratio of 0.97 and a p value of 0.05. In-hospital mortality was also assessed in two trials and showed a borderline increase in the algorithm group, with a risk ratio of 1.13 and a p value of 0.05. The interpretation given alongside this finding linked it to earlier recognition of end-of-life trajectories and greater use of comfort-focused care, rather than to harm from the intervention. The concurrent increase in do-not-resuscitate documentation was consistent with that pattern.
Family-perceived unmet palliative care needs were assessed with the NEST in two trials with 262 participants. Scores were lower in the algorithm group from day 1 to day 3 and from day 1 to day 7, but neither difference reached statistical significance. Certainty of evidence for these NEST outcomes was rated low because of concerns about bias and imprecision.
Family-reported depression, anxiety, post-traumatic stress symptoms and goal-concordant care also showed no significant differences between groups. These outcomes were each reported in 2 trials with 262 participants. The limited number of trials and participants reduced the statistical power to detect meaningful differences in these patient-centred and caregiver-reported measures.
Operational Challenges Limited Wider Effects
Four trials with 123,535 patients contributed to the pooled ICU admission analysis. No significant overall difference was observed between algorithm and usual care groups. Subgroup analysis also found no effect on ward-to-ICU transfers. In the emergency department setting, one trial reported a significant reduction in ICU admissions, but the evidence for a setting-specific benefit remained inconclusive.
No significant differences were found for 30-day hospital readmission, hospital length of stay or ICU length of stay. Two trials assessed readmission and showed a lower rate in the algorithm group, but the difference was not statistically significant. Four trials assessed hospital length of stay and two trials assessed ICU length of stay, with no significant pooled effects in either case.
Several factors were identified as influencing implementation. In patients transferred from general wards to the ICU, palliative care was often introduced alongside ongoing curative treatment, which created difficulties. One trial reported that fewer than half of eligible patients ultimately received consultations, largely because of workforce limitations. Large-scale implementation across multiple emergency departments also faced organisational challenges linked to leadership engagement, staff education and consistent buy-in across sites. Differences in algorithm parameters, the need for human input and variation in staff training may also have affected how interventions were triggered and interpreted. All included trials were conducted in the United States, which limited generalisability to other countries.
Automated EHR algorithms increased palliative care consultation rates and improved do-not-resuscitate documentation across diverse clinical settings. The strongest effects were seen in access to consultation and advance care planning documentation rather than in hospice use, ICU admission, length of stay or family-reported psychological outcomes. The interventions differed in design, setting and trigger criteria, but all aimed to identify patients with unmet palliative care needs earlier and prompt action within routine clinical workflows. Organisational factors, workforce capacity, patient acceptance and variation in algorithm design appeared to shape how these systems performed in practice. The overall pattern supports automated identification as a practical route to more timely palliative care delivery.
Source: npj Digital Medicine
Image Credit: iStock
References:
Hou CW, Hu MC, Gautama MSN et al. (2026) Automated algorithms for identifying patients requiring palliative care: a systematic review and meta‑analysis. npj Digit Med: In Press.
