A TWO-STAGE DEEP LEARNING FRAMEWORK FOR EARLY PREDICTION OF ACUTE KIDNEY INJURY USING PRE-ADMISSION KIDNEY FUNCTION TRAJECTORIES IN ELECTIVE ADMISSIONS

Acute kidney injury (AKI) affects about 13% of hospitalized patients and is linked to higher morbidity and mortality. Existing prediction models mainly target ICU populations and operate either through continuous rolling or single-time prediction at admission. We developed a novel staged deep learning model that integrates all serum creatinine (S-Cre) data from the 180 days before admission to predict AKI risk within the first three days of hospitalization in elective inpatients.

We established a retrospective cohort of adult inpatients aged 18-90 years admitted to a tertiary medical center in Taiwan between 2003 and 2021. Patients with end-stage kidney disease, nephrectomy, no S-Cre measurements during hospitalization, persistent acute kidney disease (AKD), or non-elective admissions were excluded. For each outpatient (OPT) S-Cre within 7-180 days before admission, pre-baseline S-Cre was the lowest OPT value in the preceding 180 days, and baseline S-Cre was the latest OPT value within that 7-180-day window. Baseline AKD was defined within 7-180 days before admission by a S-Cre increase ≥1.5-fold or an eGFR decrease ≥35% from the pre-baseline value, or a S-Cre increase ≥0.3 mg/dL within 48 hours. Inpatient AKI and its stages were defined according to KDIGO criteria. The cohort was split into training, testing, and validation sets (64:16:20) using 195 features, including demographics, prior inpatient AKI, laboratory data, vital signs, medications, visit frequency, and ICU status. We developed a two-stage model to predict inpatient AKI stage ≥ 2: the first stage estimated risk of AKI on the first day of hospitalization (day 1) using pre-admission data, and the second stage incorporated day 1 data to predict AKI on the next two days (days 2-3) in patients without AKI stage ≥ 2 identified on day 1. Sankey plot illustrated the data flow for each model stage. The primary performance metric was AUROC. SHapley Additive exPlanations (SHAP) plots identified the top features driving risk predictions.

Among 381,800 eligible elective admissions, the median age was 56 years, 50.9% were male, and 2.5% had baseline AKD. Median hospital stay was 5 days (IQR 3–8), with 0.4% admitted to the ICU on day 1. Hypertension, cancer, and diabetes were present in 28%, 32%, and 17%, respectively. A total of 346,137 elective inpatient visits with day1 S-Cre measurements were eligible for the first stage prediction, with 6,816 (2.0%) developed AKI stage ≥2 (Figure 1). After excluding visits with AKI stage ≥2 on day 1 or missing S-Cre on days 2-3, 319,393 visits remained for the second stage prediction, with 1,007 (0.3%) developed AKI stage ≥2. The model achieved AUROCs of 0.85 for day 1 and 0.86 for days 2-3 in predicting AKI stage ≥2 (Table 1). Key predictors for AKI stage ≥2 on day 1 included age, visit frequency, and laboratory data on liver function, kidney function, and lipid profiles. Baseline S-Cre was the leading predictor for AKI stage ≥2 on days 2-3 (Figure 2-3).

Our two-stage deep learning framework for predicting inpatient AKI performed comparably to existing models and is applicable to elective admissions, enabling risk assessment on the first day and the following two days. The next step is to develop risk-based action plans and evaluate how this AI-driven approach can support proactive, data-integrated early kidney injury prevention and management.