Nephrology · Medical Literature
Early Nephrology Consultation and Acute Kidney Injury in Hospitalized Patients: A Randomized Clinical Trial.
Abstract
To determine whether a structured early nephrology consultation triggered by a machine-learning acute kidney injury (AKI) risk score (electronic signal to prevent AKI [ESTOP-AKI]) in patients at high risk for stage 2 AKI improves patient outcomes. This randomized clinical trial was conducted at the University of Chicago, Illinois, from March 13, 2019, to August 21, 2024, in hospitalized patients with no serum creatinine (SCr)-based AKI and an ESTOP-AKI score more than 0.01. Patients were randomized to receive a structured early nephrology consultation (ENC) from an attending nephrologist or usual care (UC). The ENC included an in-person assessment and recommendations regarding volume status, kidney perfusion, medication dosing and selection, electrolytes, nutritional needs, and further testing. Patients in the UC arm only received a nephrology consultation when clinically requested by the primary team. The primary outcome was the peak change in SCr from enrollment (ΔSCr) during the 7-day follow-up. Secondary outcomes included development of AKI, need for kidney replacement therapy, and inpatient and 90-day mortality. Of the 180 patients randomized (median [IQR] age, 62.5 [50.0-71.0] years; 102 males [56.7%]), 89 (49.4%) received ENC, and 91 (50.6%) received UC. There was no significant adjusted mean (SE) difference in the 7-day ΔSCr between the ENC and UC groups, adjusted for the ESTOP risk group (0.04 [0.07] mg/dL vs -0.03 [0.07] mg/dL; P = .30), among the 70 patients (38.9%) in the development of stage 1 or higher AKI (37 [42%] vs 33 [36%]; P = .47) or among the 29 patients (16.1%) with stage 2 or higher AKI (17 [19%] vs 12 [13%]; P = .28). During the study period, there were 121 ENC consultations containing 270 recommendations compared with 19 UC consultations and 36 recommendations. Medication dosage and discontinuation, diuretics or fluids, and vasopressor recommendations were more likely to be completely followed in the UC arm (15 of 22 [68%]) compared with in the ENC arm (48 of 116 [41%]). Over the 90-day follow-up, there was no significant difference in readmission rates (ENC: 30 [34.1%] vs UC: 40 [44.4%]; P = .21) or 90-day mortality (ENC: 13 [14.8%] vs UC: 17 [18.7%]; P = .62) between the ENC and UC arms. In this randomized clinical trial of structured ENC triggered by a machine-learning AKI risk score, there was no difference in ΔSCr. AKI consultation recommendations were not followed the majority of time; whether increasing adherence to recommendations could improve outcomes deserves further study. ClinicalTrials.gov Identifier: NCT03590028.
Clinical relevance
In hospitalized patients without established AKI who are at high risk for stage 2 AKI as identified by a machine-learning risk score (ESTOP-AKI), does a structured early nephrology consultation (ENC) reduce peak serum creatinine rise over 7 days compared with usual care? __SNAPSHOT__ {"studyDesign":"Single-center, parallel-group randomized controlled trial","setting":"University of Chicago (academic medical center), Chicago, Illinois, USA; inpatient general wards and ICU; March 2019 – August 2024","sampleSize":"n = 180 (ENC: n = 89 [49.4%]; usual care: n = 91 [50.6%])","population":"Hospitalized adults without serum creatinine–defined AKI at enrollment, with a real-time machine-learning ESTOP-AKI risk score > 0.01 (indicating high predicted risk for KDIGO stage 2 AKI). Median age 62.5 years; 43.3% women; 52.8% White. Baseline characteristics were balanced between arms.","intervention":"Structured early nephrology consultation (ENC) delivered by an attending nephrologist, triggered automatically by the ESTOP-AKI score. Consultation included an in-person assessment with recommendations on volume status, kidney perfusion, medication dosing and selection, electrolyte management, nutritional needs, and diagnostic testing.","comparator":"Usual care (UC): nephrology consultation provided only when requested by the primary team. UC patients received substantially fewer consultations and recommendations.","primaryOutcome":"Peak change in serum creatinine from enrollment (ΔSCr) during the 7-day follow-up period","followUp":"7 days (primary outcome); 90 days (secondary mortality and readmission outcomes)","fundingSource":"Not reported in available abstract; no industry sponsorship disclosed in available sources"}
Key findings
PRIMARY OUTCOME: There was no statistically significant difference in peak ΔSCr between the ENC and usual care arms in adjusted analyses (ENC: +0.04 mg/dL vs. UC: −0.03 mg/dL; adjusted difference not reported as a standalone figure; P = 0.30). Confidence intervals for this comparison are not reported in the available abstract. SECONDARY OUTCOMES: Overall, 70 patients (38.9%) developed AKI of any stage during the 7-day follow-up. The incidence of KDIGO stage 1 or higher AKI did not differ significantly between groups (ENC 42% vs. UC 36%; P = 0.47), nor did the incidence of KDIGO stage 2 or higher AKI (ENC 19% vs. UC 13%; P = 0.28). There were no significant between-group differences in need for kidney replacement therapy, inpatient mortality, 90-day mortality (ENC 14.8% vs. UC 18.7%; P = 0.62), or 90-day hospital readmission (ENC 34.1% vs. UC 44.4%; P = 0.21). A numerical trend favoring ENC on readmission was observed but did not reach significance. ADHERENCE: The ENC arm generated 121 consultations with 270 recommendations vs. 19 consultations and 36 recommendations in the UC arm. Despite greater volume of recommendations, overall adherence was only 48% across both arms. Adherence was notably lower in the ENC arm (41%) compared to the UC arm (68%) for medication-related recommendations (dosing, discontinuation, diuretics/fluids, vasopressors). NNT/NNH: Not calculable from reported data given non-significant outcomes. No clinically important adverse events attributable to ENC were reported. All effect sizes above are as reported; no values were independently calculated.
Practice implications
This trial establishes that, in a single academic center, triggering structured early nephrology consultation via a validated machine-learning risk score did not reduce peak serum creatinine rise or prevent progression to KDIGO-defined AKI compared to usual care over 7 days. No benefit was observed across any prespecified secondary outcomes at 7 days or 90 days. The trial does NOT establish that machine-learning AKI risk prediction is clinically unhelpful — rather, it demonstrates that risk identification alone, coupled with advisory (non-mandatory) consultation, is insufficient to change kidney outcomes. A critical mediating factor was poor adherence: fewer than half of ENC recommendations were followed, and adherence was paradoxically lower in the intervention arm than in the usual-care arm for high-impact medication-related recommendations. Possible reasons include that pre-AKI recommendations (before marked creatinine elevation) are perceived as less urgent by primary teams accustomed to acting only after overt injury has developed. The finding that approximately 40% of identified high-risk patients developed AKI regardless of arm underscores the need for more effective prevention strategies. Findings apply specifically to hospitalized adults identified as high-risk by this particular ML score at a single academic center; generalizability to community hospitals, different patient case-mixes, ICU-only populations, or healthcare systems with different consultation culture is limited. Practice should not change on the basis of this trial alone; it does not support routine early nephrology consultation driven by ML scores without accompanying adherence-enforcement mechanisms. Only the abstract was available for this review, which limits full assessment of baseline covariate balance, subgroup analyses, statistical power calculations, and handling of missing data. __LIMITATIONS__ Single-center design at a high-volume academic hospital severely limits generalizability to community or regional settings with different consultation infrastructure and culture.,Small sample size (n = 180) raises concern about inadequate statistical power to detect modest but clinically meaningful differences; the trial may have been underpowered for secondary endpoints.,Critical adherence failure: only 48% of all recommendations were followed, and the ENC arm had lower adherence (41%) than UC (68%) for key medication recommendations, making it impossible to distinguish a true null effect of ENC from an implementation failure.,Non-mandatory, advisory consultation structure: recommendations were framed as standard consultation notes rather than binding orders, creating a structural barrier to behavior change.,The primary endpoint (peak ΔSCr over 7 days) is a surrogate marker; the trial was not powered for hard clinical endpoints (mortality, dialysis dependence, CKD progression).,Extended follow-up (90 days) secondary outcomes were likely underpowered given the sample size.,No blinding was possible given the nature of the intervention; primary teams in the ENC arm were aware of nephrology involvement, potentially influencing care in unmeasured ways.,Selection bias: enrolling only patients who triggered an ESTOP-AKI score >0.01 creates a pre-selected high-risk cohort that may not represent the full spectrum of hospitalized patients at AKI risk.,Duration of enrollment (March 2019 – August 2024) spanning the COVID-19 pandemic may have introduced temporal confounding in patient acuity, care practices, and staffing.,Only the abstract is publicly available for this review; full covariate balance, adjusted secondary outcome statistics, confidence intervals, subgroup data, and power calculation details could not be verified. __BOTTOM_LINE_CATEGORY__ not_ready_for_clinical_application __BOTTOM_LINE__ A structured early nephrology consultation triggered by a real-time machine-learning AKI risk score did not reduce peak serum creatinine rise or prevent AKI development over 7 days compared with usual care in hospitalized high-risk patients, with null results across all secondary endpoints including 90-day mortality. The trial's most consequential finding may be the low adherence to consultative recommendations (48% overall), suggesting that generating specialist recommendations before overt AKI develops is insufficient without enforcement mechanisms to ensure implementation. This study should inform future trial design — particularly around adherence-promotion strategies, pharmacist engagement, and order-set integration — rather than serve as justification for abandoning ML-guided AKI prevention programs.