A continual prediction model for inpatient acute kidney injury
Introduction
Acute kidney injury (AKI), formerly known as acute renal failure, is a sudden loss of kidney function. AKI affects 5–7% of hospitalized patients [[1], [2], [3]] and 22–57% of patients in intensive care units [[4], [5], [6], [7], [8]]. It can lead to serious medical complications and is potentially fatal. It also results in longer hospital stays and thus contributes to increasing healthcare costs [1]. Even after resolution, it can subsequently lead to severe kidney problems such as chronic kidney disease and progression to dialysis dependency. The incidence of AKI is highest in elderly patients [9,10], and its rate has been steadily increasing in this population due to an increasing number of comorbidities, aggressive medical treatments, and greater use of nephrotoxic drugs. Two factors complicate AKI diagnosis: it has a heterogeneous etiology, and it often develops stealthily in hospitalized patients being treated for other problems. However, up to 30% of hospital-acquired AKI is preventable [11] if predicted in time. AKI is also potentially reversible if diagnosed and managed in time.
The seriousness of AKI and its preventability make AKI a perfect candidate for predictive analytics. Hence many machine learning based predictive models have been built to predict inpatient AKI from electronic health records (EHR) data; good reviews of these models can be found in Refs. [12,13]. All these models had a particular time when AKI prediction was made for the rest of the hospital stay, for example, at 24 h after admission [14,15], or at 48 h after admission [16,17], or around the time of a medical intervention [[18], [19], [20]]. However, there are two fundamental problems with such models that have a fixed time for predicting AKI. First, hospital stays can be several days long, during which a patient's medical condition can significantly change and after which AKI may develop within a few hours. For example, in our data, 39.4% of AKI incidences occurred after five days from admission and 15.7% occurred after 10 days from admission. Therefore, it is difficult to predict such incidences too far ahead in time. Second, if a prediction model has a fixed time of prediction, for example, 24 h after admission, then it is bound to miss all the AKI incidences that occur during those 24 h. For example, in our data, 12.8% of AKI incidences occurred within 24 h from admission and 30.9% of AKI incidences occurred within 48 h from admission. Therefore, the later the prediction time, the more incidences it will naturally miss.
Since a prediction time that is neither too early nor too late is desirable, AKI should be predicted continually during the entire inpatient stay so as to optimally predict it before it develops. This possibility was recently mentioned in a workgroup statement from the 15th Acute Dialysis Quality Initiative (ADQI) consensus conference [12] as a model that would generate a prediction score in real time as each new data value is received. However, to the best of our knowledge, no previous work has reported such a model to continually predict AKI. Although susceptible patients are continually monitored manually, no automated models have been reported that can continually monitor patients’ variables in EHR to predict AKI. Automated continual prediction can not only reduce the manual work of caretakers, but is also particularly desirable given that AKI can stealthily develop in hospitalized patients being treated for other problems.
Our contributions in this paper are as follows. We introduce a novel framework for automatically predicting AKI continually over hospital stays. A trained machine learning model is used to predict AKI every time a patient's status changes in the EHR. For example, it will predict AKI every time a new medication is prescribed, or a new comorbidity is recorded, or a new laboratory test result becomes available. This new framework does not require constant monitoring and is designed to work automatically using EHR data and trigger an alarm when the potential for AKI increases. We also introduce an evaluation method that comprehensively measures the overall performance of such dynamic predictions. Furthermore, we retrospectively measure the performance of advance predictions, that is, how well the model could predict AKI, say, 6 h in advance, or 12 h in advance etc., which was not measured in previous studies. Our continual prediction framework also lends itself to discovery of the most important dynamic features in which change in values often trigger the prediction of AKI. To the best of our knowledge, researchers in previous work had not reported the dynamic importance of features in predicting AKI.
Section snippets
Related work
The only work we know that is close to our work is by He et al. [21] in which they present a framework in which prediction is made after every 24 h for AKI to occur within the next 24 h. They set time for AKI incidence as the day on which AKI is diagnosed. Thus their framework has a granularity of 24 h. In contrast, our framework has no granularity restriction and it continually predicts AKI to occur any time during the rest of the hospital stay. Since AKI can develop within a few hours
Data collection
The data was collected from the EHRs of 15 hospitals that are part of Aurora Health Care system. These hospitals are located in southeastern region of Wisconsin state and use the same EHR system. Structured data along with their timestamps for entire hospital stays was collected for all adults older than 60 in 2013, 2014, and 2015 (number of patients = 84,480). We focused on hospitalized older adults because the occurrence of AKI is especially common in this age group [9,10], otherwise our
Continual vs. one-time prediction models
We first compare our proposed continual prediction model with the one-time-at-24-h prediction model and then with many other one-time prediction models. Table 3 shows a comparison between AUC obtained by the continual prediction model and by the one-time-at-24-h prediction model through 10-fold cross-validation using exactly the same folds. The first row shows results in which we exclude hospital stays in which AKI developed within 24 h from admission (483 such hospital stays), because it is
Conclusions
A new framework of continual prediction from EHR data was introduced and applied for predicting AKI. Instead of applying the trained model at a particular chosen time, as has been done in the past, the continual prediction model was applied continually over entire hospital stay whenever any patient variable changed. A method to evaluate the overall performance of a continual prediction model was also presented. Our experiments on a large dataset showed that the model out-performed one-time
Declaration of competing interest
The authors declare no competing interests.
References (44)
- et al.
Hospital-acquired renal insufficiency: a prospective study
Am. J. Med.
(1983) - et al.
Hospital-acquired renal insufficiency
Am. J. Kidney Dis.
(2002) Acute kidney injury in elderly persons
Am. J. Kidney Dis.
(2010)- et al.
Physician prevention of acute kidney injury
Am. J. Med.
(2015) - et al.
Predictive models for acute kidney injury following cardiac surgery
Am. J. Kidney Dis.
(2012) - et al.
Modeling disease progression via multi-task learning
Neuroimage
(2013) - et al.
KDOQI US commentary on the 2012 KDIGO clinical practice guideline for acute kidney injury
Am. J. Kidney Dis.
(2013) - et al.
Risk factors for acute kidney injury in older adults with critical illness: a retrospective cohort study
Am. J. Kidney Dis.
(2015) - et al.
What can natural language processing do for clinical decision support?
J. Biomed. Inform.
(2009) - et al.
Acute kidney injury, mortality, length of stay, and costs in hospitalized patients
J. Am. Soc. Nephrol.
(2005)