Acute kidney injury: a risk scoring system for general surgical patients

Background

It has been suggested that worldwide up to 20% of people admitted to hospital either have or will develop AKI (Susantitaphong et al, 2013). Although the incidence in the UK may be lower, it can have high cost implications for both the patient and the hospital (Thomas et al, 2015). It is estimated that AKI in the UK is associated with 100 000 deaths a year (National Confidential Enquiry into Patient Outcome and Death (NCEPOD), 2009). In 2013, NICE estimated that the cost to the NHS was in the range of £434-620 million a year (NICE, 2013b).

Given the significant cost implications of AKI for the NHS, various plans have been developed to identify and manage patients with AKI. In the 2015/16 Commissioning for Quality and Innovation (CQUIN) goals, one of the new indicators added was the care of patients with AKI. The aim was to improve the follow-up care for these individuals and reduce re-admission (NHS England, 2015). The rationale was based on studies suggesting that the economic annual cost of AKI-related inpatient care was escalating and the cost to the NHS was in excess of £1 billion (Kerr et al, 2014). It has been observed by Ozrazgat-Baslanti et al (2016) that AKI can lead to the development of chronic renal failure and increased mortality.

These goals followed on from a safety alert issued in 2014 by NHS England, with the aim of standardising the early identification of AKI in patients in all hospitals providing pathology services. The alert provided an algorithm that allowed timely and consistent detection and diagnosis of AKI in patients (NHS England, 2014). Although this algorithm allowed the laboratory to alert clinicians to the occurrence of AKI, it did not help with its prevention.

NICE (2013a) and NCEPOD (2009) suggested that ideally AKI should never occur. NICE (2013a) made the point that identifying patients at risk of AKI is crucial to its prevention. If early preventive measures can be implemented, then nursing staff can be made aware of at-risk patients, applying care plans or strategies to help reduce a patient's likelihood of developing AKI during their hospital stay. One method of detecting these at-risk patients is by using risk assessment tools or prediction models. Prediction models are developed to estimate the probability of risk that a specific event will occur (Collins et al, 2015). They help medical and nursing staff make appropriate clinical decisions and allow patients to make an informed choice about their treatment (Pavlou et al, 2015). Regarding patients at risk of AKI, risk prediction tools have been used and are established in some clinical settings, such as cardiac medicine, but less so in the general or acutely unwell population (NICE, 2013a). Harris et al (2015) reported that, due to the interaction of surgical events and comorbidities causing renal dysfunction, surgical patients have unique risk factors that can lead to the development of AKI.

A search of the literature was conducted in October 2017 on the databases Medline and Cumulative Index to Nursing and Allied Health Literature (CINAHL) using the key words ‘acute kidney injury’ and ‘prediction model’. Limiters were added to only include texts that had full text availability and that were published between 2012 and 2017. All studies examined confirmed that the prediction of AKI was important and reinforced that many such events were avoidable.

All studies showed a lack of a model based on surgical patients. All the studies reviewed were based on models using data from medical patients. As reported by Hodgson et al (2017a), caution should be applied when using models across specialties, so this confirmed the need for a model to be developed specifically for surgical patients. The literature review concluded with a list of common variables suggested by the articles that can influence the risk of developing AKI. These variables included age, chronic kidney disease, hypotension, sepsis, diabetes, heart disease/failure, liver disease and pancreatitis. They were used in the statistical analysis and development of a prediction model.

In order to develop a method of predicting the risk of the development of AKI in general surgical patients, Hospital Episode Statistic (HES) coded data (NHS Digital, 2019a) was used to determine factors likely to contribute to its development and devise an admission checklist that would identify those patients at greatest risk.

Aim

The aim of the study was to develop a prediction model that could be used as the basis for an admission checklist for surgical patients, giving each patient admitted a risk score for developing AKI during their hospital stay. If a patient is found to be at high risk, this would alert staff and a plan of care could be implemented to reduce the risk of AKI occurring.

Development of the model

From a literature review, the author's experience and discussions with medical and nursing staff, it was possible to produce a list of 59 potential coded variables that may have contributed to the risk of AKI. In addition to these 59 variables, 2 more were added: age over 60 years, based on a study by Kate et al (2016), and a coded diagnosis of shock, based on a study conducted by Suarez-de-la-Rica et al (2017). Within the 61 potential risk factors, the related ones were grouped, aggregated into 25 groups (Table 1).

Anonymised HES general surgical data from two randomly selected trusts with a full range of general surgical specialties were obtained for a 2-year period (October 2015 to October 2017). The data, provided by the hospital trusts, included the following parameters: age, sex, and coded information showing diagnoses (ICD10) and procedures (OPCS). In total, 19 163 patients' consecutive admission data was used for analysis.

At this stage, four variable groups were excluded due to their low incidence (fewer than 25 patients with these codes found in the HES data).

With assistance from CRAB Clinical Informatics Ltd and the medical module of its CRAB (Copeland's Risk Adjusted Barometer) system, which allows the transformation of clinical ICD10 diagnostic codes into potential harm events (trigger events), it was possible to examine the relationship between potential risk factors and the incidence of AKI. The trigger events were based on the UK version of the Global Trigger Tool (Adler et al, 2008; Classen et al, 2008). These trigger events built within the medical CRAB system identify instances of potential harm to a patient. One of these trigger events is the development of AKI. Of the 19 163 patient admissions, a total of 589 patients with AKI were identified (3.1%).

Using the statistical and mathematical analytical program Analyse-it (Analyse-it Software Ltd), it was possible to create a binary logistic model using yes/no answers to indicate whether the patient had any of these variable groups and correlate this with the determinant variable of AKI. This created a logistic regression equation (In Odds ratio = β0 + β1x1 + β2x2 …). From the first run of the logistic regression equation it was possible to identify the most significant variables that contributed to an increased incidence of AKI. Any variable group with a statistical significance predicting AKI of P>0.05 were excluded and the analysis was rerun. The second run included 13 variable groups, 8 groups having been excluded. These groups were liver cancer, dementia, stroke and effects of a stroke, hypertension, atrial fibrillation and flutter, appendicitis, chronic liver disease and chronic obstructive pulmonary disease.

On the second analysis, all of the constant weightings were statistically significant (P<0.05). The logistic regression equation showed a statistical significance of P<0.001. After the second logistic regression equation was created, it was necessary to identify the value of all the constants to highlight the patients at the lowest risk and the highest risk of AKI. The weightings for each variable group are illustrated in Table 2.

Scoring system and checklist

From the weighted variable groups it was possible to create a simple scoring technique that would allow easy identification of those patients at higher risk of AKI. From this second logistic regression equation it was possible to transform the constant weighting into a numerical value. This created a list of variable groups that each had a score of 1, 2, 4 or 8 depending on the risk of AKI. Anyone with an overall score of 0-1 had a 0-2% risk of developing AKI, a score of 2-5 gave a risk of 2-10%, a score of 6-9 a risk of 10-45%, a score of 10-13 a risk of 45-85% and any score over 14 had an 85% risk of developing AKI.

The scoring system was then used to create the admission checklist. See Figure 1 for the completed checklist. The checklist will be used on admission to the surgical ward.

Sensitivity and specificity check

The sensitivity and specificity of the model was tested by the application of the receiver operating characteristic (ROC) curve (Figure 2). Results showed a good correlation between the score and the occurrence of AKI across the whole risk spectrum with the area under ROC being 0.806 (95% confidence intervals between 0.787 and 0.825).

Discussion

There can be little doubt from the literature that AKI is a common adverse event following hospital admission and is likely to increase in frequency, given the ageing population (Han et al, 2016). Several sources have suggested that the cost implications of AKI are significant and that considerable savings may be achieved by eliminating its occurrence. Although NICE (2013b) suggested that the cost to the NHS may be as high £434-620 million a year, Kerr et al (2014) estimated the cost to be in excess of £1 billion. Estimating the cost implications of any adverse event can be difficult and many models take into account the increase in length of stay resulting from the development and treatment of AKI. In the context of the NHS, where hospital beds are at a premium and occupancy levels often approach 100% in acute care facilities, the figure from Kerr et al may be an overestimate. It is likely that the NICE estimate is closer to the truth, particularly as not all cases of AKI are avoidable, being a consequence of the disease process itself, despite all efforts to prevent it. Whatever the cost implications, the impact on the patients themselves is significant and AKI can have serious and significant long-term sequelae.

There have been a number of attempts to define AKI, leaving some researchers to argue that an ideal definition may not actually exist (Thomas et al, 2015). However, the definition that this article has used is the KDIGO (2012) definition. It is the most recent definition to date, the one used by most hospitals in the UK and the one used by coding departments to record the occurrence of AKI.

In order to assess the factors that may have an influence on the occurrence of AKI, an advanced search of the literature was conducted in October 2017. Other potential factors were added to this list based on personal observations and discussions with medical and nursing colleagues. Many articles had focused on general medical patients and few on general surgical patients. However, the variables selected were compatible with both general medical and surgical articles previously published and with the experiences of medical and nursing colleagues.

Although a number of authors had attempted to produce models that would predict the risk of developing AKI (Kate et al, 2016; Hodgson et al, 2017a; 2017b; Ohnuma and Uchino, 2017) these were in medical specialties rather than general surgery. Hodgson et al (2017a) examined a number of prediction models and suggested that, as the disease processes are particular to individual specialties, a global prediction tool would be difficult to produce and should not be applied across specialties.

HES data provides information about patients admitted to UK hospitals. Coding is performed chronologically on discharge and is divided into episodes of care (NHS Digital, 2019a). In general medicine there may be many episodes of care and the ultimate diagnosis may not be made until later episodes. Some methodological software systems—Summary Hospital level Mortality Indicator (SHMI) and Hospital Standardized Mortality Ratio (HMSR)—only examine the first few episodes of care and may therefore not indicate all relevant diagnoses (Bottle et al, 2011; NHS Digital, 2019b). In contrast, CRAB examines all diagnostic codes and all episodes of care and then transforms the codes into trigger adverse events that have been well validated. The CRAB medical module was therefore the ideal choice for this study (CRAB Clinical Informatics, 2019).

Logistic regression methodologies are the ideal mathematical techniques as they avoid the inaccuracies associated with linear models. The method used here assessed all selected factors and then reduced the number by removing non-significant factors until only significant factors remained. The subsequent logistic regression equation would be difficult to use practically as an assessment tool. In the future, however, once electronic medical records become more accessible and reliable, it should be possible to develop an application to produce an individualised patient prediction for the risk of AKI. In a ward setting, particularly in an emergency situation, a checklist that immediately assesses the patient and produces a rapid indication as to whether the patient is high or low risk would seem to be the ideal tool. The present design of the checklist allows a rapid assessment of the patient's risk of AKI and provides some guidance as to their management.

The checklist has been designed particularly for the general surgical patient and, as can be seen from the ROC curve, provides an extremely accurate assessment of the risk of developing AKI that can be applied across the risk spectrum.

In current hospital practice there is more focus on the identification of AKI, with the use of laboratory algorithms and laboratory alert systems rather than on the actual prevention of AKI occurring. The study here has produced a simple, accurate, easy-to-use and practical checklist, which can be used by nursing and medical staff alike to assess those patients at risk of developing AKI and provide guidance on management. The hope is that once the checklist is in common use it will reduce the incidence of AKI and thereby improve patient safety. If this occurs, although the logistic regression algorithm may require updating over time, the actual variable scores would remain unchanged even though the overall risk may change and hopefully decrease.

Conclusion

Acute kidney injury is a common adverse event following admission to hospital and in many cases is avoidable. However, in the general surgical setting there is currently more focus on the detection and subsequent treatment of AKI. Although this is laudable, from a patient's perspective, prevention is better than a cure. The current study has produced a simple, accurate, easy-to-use and practical checklist that can be used by nursing and medical staff to identify those at risk in order to instigate preventive measures.

KEY POINTS

CPD reflective questions