Acute Kidney Injury is a relevant event that pervades health care systems and has poor outcome. Although the burden of AKI varies depending on its classification, on the use of administrative or clinical data , on the clinical setting and on differences between high and low resource countries, an increase in AKI episodes has been observed over the last two decades: a study conducted in England between 1998 and 2013 reported that hospital-acquired AKI not requiring dialysis increased from 317 to 3995 cases per million population (pmp) . In the cases of AKI requiring dialysis (AKI-D), a retrospective study conducted in the USA demonstrated an increase in episodes from 222 to 533 pmp between 2000 and 2009 . In the ICU, critically ill patient AKI incidences varied from 10 to 50% [18,19,20,21].
AKI is linked to a high mortality rate, ranging from 10 to 60%, depending on its severity and the concomitant failure of other organs. Surviving patients may develop decreased renal function, need for chronic dialysis treatments and an increased risk of cardiovascular disease and subsequent mortality .
According to the AKIN guidelines, a serum creatinine increase (AKI-sCr), “or” a urine output decrease (AKI-Uo), defines the presence and relevance of AKI . When urine output is considered, whether or not it is associated with an increase in serum creatinine, a higher number of AKI cases are diagnosed, and the detection can be made earlier . The duration of oliguria appears to be associated with the beginning of dialysis and an increased risk of death . The definition of oliguria has recently received critical appraisal: the impact of both the volume of urine as well as the duration of oliguria on AKI prediction is controversial: KDIGO Clinical Practice Guidelines for Acute Kidney Injury  define a urine output < 0.5 ml/kg/h for > 6 h as AKI stage 1, < 0.5 ml/kg/h for > 12 h as AKI stage 2, and a < 0.3 ml/kg/h for 24 h or anuria for 12 h as AKI stage 3.
Prowle et al. reported that only 15% of ICU patients with an episode of oliguria developed AKI-sCr stage 2. If oliguria persisted for at least 12 h, the relative risk of developing AKI-sCr the following day was equal to 11.5 with a positive likelihood ratio of 13.5 . Macedo et al. observed oliguria before serum creatinine increase, and this allowed an earlier diagnosis of AKI stages 2 and 3: 6 consecutive hours with UO < 0.5 ml/kg/h were linked to the highest rate of progression to AKIN-sCr stage 2 (79%). A urine volume of less than 0.72 ml/kg/h for 24 h was able to predict AKIN-sCr stage 3 with positive and negative values of 0.37 and 0.76 . Ralib et al. observed that a 6-h urine output threshold of 0.3 ml/kg/hour best-predicted mortality and need for dialysis, with a positive and negative predictive value of 0.34 and 0.90, instead of 0.28 and 0.89 as reported for serum creatinine . The prospective FINNAKI study  reported an increase in AKI-sCr risk for a 3–6 h period of urine output < 0.1 ml/kg/h and an increased risk of mortality at 90 days.
We investigated the predictive value of urine output on AKI stage 2/3 using two large databases of patients admitted to ICUs, the eICU and the MIMIC-III [9, 10].
In the case of “big data”, the information can be elaborated mathematically with analytic methodologies. With regard to medicine, the use of artificial intelligence has been significantly developed in two branches: physical and virtual. The virtual branch consists of machine learning methods, characterized by algorithms and statistical models that learn from data that are able to recognize and deduce patterns.
Our study applies a machine learning method to the determination of the risk of AKI development considering urine output, and consequently evaluating the algorithm to define an electronic alert. Here, two methods were used and compared: the deep learning model obtained the highest sensitivity and specificity of 82% with an AUC of 0.89. The accuracy of this model was confirmed with the high positive and small negative likelihood ratio. The performance of deep learning appears to be superior to the logistic regression model. Logistic regression reveals a likelihood positive ratio of 3.13 and a likelihood negative ratio of 0.32, with a urine output threshold of 0.37 ml/kg/h for 7 h.
The deep learning model differs from the previous method as it does not consider urine output in terms of volume and time, but it analyzes the dynamic changes: the result is the higher accuracy of the predictive score, with the highest positive likelihood ratio equal to 5 and the lowest negative likelihood ratio equal to 0.2.
The deep learning model was able to predict 88% of AKI cases at least 12 h before the event: for every 6 patients identified as being at risk of AKI by the deep learning model, 5 experienced the event. For every 12 patients not considered to be at risk by the model, 2 developed AKI.
AI can analyze the relationship between “big data” as “input” and “events” as “output”.
Concerning AKI prediction in the ICU, the study of Huang et al. analyzed 9,791 MIMIC III and eICU patients  and investigated the predictive value of 52 clinical sets of data collected over 6 h. The data included 8 classes, such as fluid balance, demographic, anamnestic, clinical interventions, and laboratory results. The AdaBoost predictive algorithm revealed the highest AUC = 0.88 for AKI-sCr and/or AKI-Uo occurrence during the first week of an ICU stay.
Flechet et al.  analyzed the performance of the prediction model for the development of AKI in the ICU by using data from the retrospective study of EPaNIC. Data used for the prediction models were based on information obtained at ICU admission and after the first 24 h, and also included the total amount of urine and the slope. The outcome was AKI at any stage or AKI at stages 2 or 3, which manifested during the first week of an ICU stay. The predictive performance was also compared with mathematical models with a biochemical marker such as serum neutrophil gelatinase-associated lipocalin (NGAL) levels. The performance of the mathematical model was high, with an AUC of 0.84, which was higher than NGAL with an AUC of 0.74.
Zimmermann et al.  obtained different results using data from MIMIC-III. The input data regarded variables recorded during the first day of an ICU admission including gender, age, heart rate, blood pressure, SpO2, lab values and the hourly rate of urine output. The outcome was AKI at any stage on days 2 and 3 of an ICU stay. Univariate analysis did not show a significant association between urine output and continuous creatinine outcome.
These studies differ from our investigation because of the endpoint. AKI development is similar but the time is different: the first week of the ICU stay in the study of Huang and Flechet, and days 2 ad 3 in the study of Zimmermann. In our study, urine output is entered as continuous data during the entire ICU stay, (as the output AKI stages 2 or 3).
Our study can be applied to a generation of electronic Acute Kidney Injury alert systems (eAKI) which enable earlier detection of AKI. Some warning systems , but not all , have proven to be of benefit to the patients in terms of diagnostic procedures, treatment, and nephrology consultation . Park observed that eAKI permitted a reduction in the severe progression of renal failure and an improvement in AKI recovery, but failed to report the effects on mortality . Prendecki reported a better outcome in terms of survival and need for renal replacement therapy .
Our study triggers the calculation of a risk score obtained with continuous input of data, urine output measured as ml/kg/h, with the ability to “alert” the medical staff 12 h before the onset of AKI.
The impact of electronic alerting of AKI essentially results in the possibility of earlier medical intervention from the diagnosis of changes in clinical conditions which could be responsible for decreased kidney function (for example, dehydration or left ventricular heart failure). This intervention could lead to earlier personalized treatment . Electronic alerts for AKI can be made based on the detection of changes of a biomarker, such as serum creatinine, with the limit of basal value determinations; otherwise, eAKI can be organized on multiple markers, such as clinical, anamnestic, and biochemical markers, moreover, the implementation of a care bundle alarm can result in an improvement in outcomes . The ELAIA-1 trial, which aims to ascertain the usefulness of eAKI on disease progression and mortality, is still ongoing .
In our study the score is calculated with an algorithm that takes into account “blocks” of urine-output/hour in a dynamic movement. The result can be summarized as “for every six triggered alarms only one is false”, and “for every 12 non-triggered alarms only 2 cases are false negatives”.
Several limitations of our study should however be mentioned.
First, this is a retrospective study based on data from hospitals in the United States.
Second, the precision of an hourly urine output report could be limited, therefore automated electronic recording appears to be required.
Third, the definition of AKI based on serum creatinine may be limited due to the nadir recognition, which in this study was the lowest recorded value during the hospital stay, and a missed AKI diagnosis could occur in this scenario. We report an incidence of AKI stage 2/3 in 3% of ICU patients, and consequently the question of selection arises. Although the incidence in the ICU varies with different definitions, based on the international FINNAKI and AKI-EPI studies (38–21), 35% to 60%, of ICU patients may be affected [38, 39]. Recently, the prospective observational study of Wiersema et al. , which defined AKI according to the KDIGO criteria, reported an occurrence of stage 2/3 between 3.3 and 4.5%. Uchino observed AKI in 5.7% of 29,269 critically ill patients (41). Therefore, the incidence of AKI in our population appears to be in agreement with other studies, particularly since patients undergoing hemodialysis (AKI-D) were excluded. The decision to exclude AKI-D was based on the fact that, in our opinion, the database we used did not provide sufficient information on hemodialysis.
In conclusion, urine output can correctly predict the development of oliguric Acute Kidney Injury, and the highest accuracy was obtained with a deep learning model. The characteristics of urine output in terms of dynamic flow analysis, more than a fixed volume and time, are necessary to implement the predictive score. Its applications into an e-Alert system, which is able to automatically inform the medical staff of the risk, could be useful for reducing the incidence of AKI.