Acute kidney injury (AKI) develops in 4–10% of hospitalized patients, of which 35–40% are hospital-acquired. Patients with AKI experience prolonged hospitalization which subjects them to more nosocomial infections. More recent evidence demonstrates the high risk of death and kidney failure over 3–5 years following AKI. Urinary biomarkers of early renal tubular injury can predict AKI, but the timing of biomarker assessment is uncertain since the onset of AKI is highly variable. It is also not cost-effective to perform repeated and frequent AKI biomarker testing in all patients. The advent of electronic health records allows for the integration of multiple patient-data points that are routinely available. Conventional AKI definitions based on serial serum creatinine are numerical and easily configured as a binary outcome for predictive analytics. We applied supervised machine learning to a historical inpatient database of laboratory parameters belonging to 20,732 cases and predicted hospital-acquired AKI with AUC above 0.8 in the derivation, training, and testing cohorts. We validated this algorithm in a more current 40,332 inpatient database, and incorporated electronic medication records as features, with an improved AUC of 0.84 for AKI prediction. Clinical informatics are not without their challenges. Firstly, electronic data acquisition, integration, cleaning, and normalization, are labour and resource-intensive upfront. The electronic AKI definition does not naturally exclude cases with baseline kidney failure on hemodialysis and efforts are made to curate the cohort of interest. It is unclear what is the desired lead-time for optimal AKI prediction. A prediction window under 24 hours allows for superior diagnostic accuracy but is too proximate to AKI; we chose an extended 48-hour prediction window that allows more time for clinically meaningful AKI prevention but compromised the model performance. We also incorporated only features with an objective time stamp to be certain of their temporal sequence in AKI development. We adopted a novel time-invariant and time-variant module to enhance the bidirectional recurrent neural network model in analyzing these time-sensitive and interactive features. Our model’s probability threshold to define predicted AKI can be varied based on clinical indications. At 15% prediction threshold, our model generated 2 false positives for every true AKI, but only predicted 26% of AKIs. A lowered 5% prediction threshold improved the recall to 60% but generated near 4,000 AKI alerts with 6 false positives for every true AKI. Representative interpretation results provided insights into the top-ranked features that predicted AKI, categorized in association with sepsis, acute coronary syndrome, nephrotoxicity, or multiorgan injury, specific to every case at risk. These solutions can be built into an artificial intelligence platform that runs synchronous to electronic health records, and generate rolling AKI predictions in real-time that risk-stratify inpatients for more selective clinical or biomarker-based AKI adjudication. The successful implementation of these advances in health informatics, however, still requires clinical protocols and medical education to be developed in managing predicted or subclinical AKI, before AKI management can truly be made pre-emptive and less reactive than our current practice.