Variation in and risk factors for paediatric inpatient all-cause mortality in a low income setting: data from an emerging clinical information network

It is important for a health system to have an accsurate picture of overall (crude) and cause-specific hospital mortality. Although using mortality or risk adjusted mortality as an indicator of quality of care is contested because it is hard to adjust for case-mix or the severity of illness on arrival (case-severity) [1‐4] the presence of 3.variation in mortality may point to possible inequalities in population health, access or resource provision that can be addressed. Yet, little attention has been paid to understanding and exploring hospital mortality and its variability in African settings, perhaps because routinely reported data are often of poor quality [5].

In this report our aim is to contribute to efforts to understand health system performance and describe mortality and its variability. We also explore whether simple clinical characteristics used to classify severity of common childhood illness are consistently associated with inpatient mortality. Demonstrating the latter provides a rationale for reinforcing their widespread clinical use and for considering them as components of a common data framework for paediatric admissions. A common data framework (potentially included in emerging electronic record systems) could improve our ability to characterise hospitals by their case-mix and case severity and inform health system strengthening efforts in support of universal access to quality health care.

A total of 44,314 children were admitted into the CIN hospitals from September 2013 to March 2015, of these 33,741 (76%) were aged 2–59 months, had outcome data and a plausible date of admission or death. Characteristics of children by hospital and overall are presented in Table 2 . The overall cluster adjusted crude mortality across hospitals was 6.2% (95% CI 4.9 to 7.8; range 2.1% - 11.0%) with five-fold variation across hospitals while the risk-adjusted mortality rate derived from Model 3 was 6.2% (95% CI 4.7 to 7.6; range 3.0% - 9.4%) (Fig. 3a). Hospital identity was significantly associated with mortality (likelihood ration test, LRT <0.001 when compared to the null model). The distribution of risk factors per admission varied by hospital and outcome with 85% of children who survived having 3 or fewer risk factors while of those who died 53% had more than 3 risk factors (see Fig. 3b).

The amount of missing data varied by variable and hospital (see Table 2 and panel 2). Pooled estimates of association weighted by proportion across missingness patterns were similar to those from model 3 (imputed data assuming MAR) and are presented in Appendix. These findings provide support for assuming that data were missing at random.

All-cause mortality is highly variable across only 12 hospitals even within a common age group. Variation in mortality was associated with the proportion of children with multiple risk factors, something that cannot be determined using data from existing routine health information systems. This variation in risk factors at presentation might be linked to the varying number of comorbidities at presentation resulting from overlapping syndromic diagnoses. This finding of major variation in the risk profile of children admitted to hospitals is important but rarely highlighted in existing research literature from low-income settings. The variation of mortality with hospital identity is probably explained by associated variability in underlying risk factors (such as malaria and HIV prevalence, socio-economic status, nutrition and access) that influence case-mix and case-severity. Although, these data exist at a population level, adjusting for these parameters would require a large number of hospitals in more defined, smaller regions. As such we have refrained from adjusting for these parameters in our models due to the risk of ecological fallacy where population estimates do not necessarily apply to the population seeking care. In addition, care seeking patterns in Kenya vary across populations and individuals with some of the patients accessing care directly from hospital while others are through referral from primary health care services but these data are rarely documented and hence difficult to disentangle within this population. Thus, although there is a standard policy guiding the provision of PHC in Kenya and how patients might progress through the system anecdotal evidence suggests procedures are rarely followed in practice and there are limited data from primary care on access and care seeking behaviors.

Mortality may also potentially be influenced by differing availability of resources across hospitals and variation in care practices shown to exist in other studies [27, 28]. More comprehensive datasets from a larger number of hospitals, as are being used in high income settings [29, 30], would however be required to try and determine the degree to which quality of care and local context are associated with inpatient mortality in low-income settings. Even then the use of mortality to infer anything about quality of care specifically is contested [1, 31]. What is more pertinent to low-income contexts is that such variation in mortality should prompt thinking on where system strengthening efforts may be most needed.

In the multivariable analyses, we demonstrate that majority of clinical risk factors included in best-practice guidelines were associated with mortality irrespective of diagnosis. This approach is important due to the inability to confirm diagnoses in these settings. For example, we have previously reported that basic investigations like blood culture, CSF analysis and urine culture are not routinely available [28] and there is no access to tests of inflammatory markers, for biochemical derangements or for detection of other pathogens except for malaria and HIV. Thus, an approach that is agnostic of diagnosis may be the more useful approach at scale. Prior demonstrations of association are typically from single sites or focus on specific diseases [15, 16, 19, 32]. We believe this is the first report of the value of these clinical risk factors from multiple settings, across all cause admissions, with routine observations made by large numbers of clinicians. As malaria prevalence and other attributes of the hospitals’ setting may be correlated, including for example proportion living in poverty and hospital size, we cannot attribute the influence of malaria prevalence on the behaviour of risk factors to malaria as a disease. However, findings suggest that future efforts to explore variability in mortality or develop risk prediction approaches should take the prevalence of malaria into account.

Our data support the practical, day to day use of these clinical factors in identifying children who may be prioritised for attention, specific treatments and review. Integrating such clinical factors into a low-income setting, patient-level prognostic score might be possible [16] replicating approaches in high-income settings [33]. However, the implementation of such scores in routine settings with limited human resources, high staff turnover and without computer assisted decision support would likely be very challenging as even basic job aides are somewhat slowly adopted [28]. These clinical features might however be used to characterise risk profiles of hospital populations aged 2 to 59 months. This might enable improved understanding of changes in mortality over time within sites using methods such as cumulative sum control charts (CUSUM) [34‐37] where risk-adjustment may facilitate exploration of variation in a single institution’s performance [1, 3, 31]. More pragmatically such risk profiles may help point to different health system challenges. High prevalence of cases with high risk factor density may suggest problems of access, late detection or delayed referral. They may also point to particularly vulnerable populations within catchment areas, where malnutrition and poverty are prevalent for example, or demonstrate the impact of varying disease ecology that should be taken into account in resource allocation. Our findings suggest that existing routine health information systems data (such as DHIS2) [38] would be inadequate for this purpose, suggesting value in developing and implementing suitable common data frameworks.

The data we report need to be interpreted in the light of their limitations. First, is the missing data problem commonly associated with collecting data in routine settings. The Clinical Information Network worked with hospitals to promote data quality that improved over time but missingness varied across variables and hospitals. Thus, just less than 60% of cases were included in our complete case multivariable models. We used multiple imputation to allow use of all available data and undertook sensitivity analyses that suggest our findings are generally robust. An alternative approach would be using Expectation-Maximization (EM) algorithms to get maximum likelihood estimates [39]. Second, our sample of hospitals is small and non-random, with selection based on feasibility and efforts to represent diverse but not all epidemiological and socio-cultural contexts. One hospital that failed to provide reasonable quality data was excluded from these analyses. As such, due to the limited number of hospitals, our models did not sufficient power to explore potentially important factors at the hospital level. Thirdly, diagnoses are clinical and rarely informed by diagnostics while risk factors such as hypoglycaemia, hypoxemia or an individual’s HIV status could not be examined as these are rarely routinely evaluated. Fourth, we did not include more robust nutritional indicators like mid-upper arm circumference or weight-for-height z scores or account for vaccination status which may influence mortality because such data are largely missing. However, there is evidence that vaccination coverage is high in Kenya and we were able to use weight-for-age z scores for nutritional assessment.

In summary, all-cause crude and cluster adjusted mortality rate was highly variable across hospitals. Such variation is largely explained by variation in severity of illness at the time of clinical presentation, findings that point to underlying differences in population health and health system performance that will need to be explored. Our data supports the use of clinical risk factors drawn from guidelines in day-to-day use in prioritizing care and identifying children at the highest risk of death but also to develop risk adjusted mortality estimates across hospitals. We also demonstrate how having a large patient level dataset from multiple geographically diverse sites may improve our understanding of health system challenges and performance. Such work provides a learning platform for the design of common data frameworks that are relevant to clinical practice and might be incorporated into future electronic medical records (EMRs) that go beyond a focus on cost-accounting needs [40]. To maximise the future value of EMRs there is a clear imperative for researchers, clinicians, policy makers and health care managers to engage in their design so they enable health system performance monitoring at scale as is occurring in specific fields such as HIV care [41, 42].