Developing a predictive model for shift-level nurse staffing using routine data - WER@INSEL project (Workforce Effectiveness Research at the Inselspital)

One substantial challenge facing healthcare systems worldwide is the aging population. Here in Switzerland, the 2017 census showed that almost 20% of the population was aged 65 or older. Accompanying rates of chronic and comorbid diseases are leading to increased health care demand. In terms of quality of care, the Swiss healthcare system is consistently ranked in the top five of over 195 countries; however, criticism is often raised concerning its costs (ranked behind only the United States). As elsewhere, then, cost containment measures have been implemented.
Switzerland’s acute care hospitals currently employ around 100,000 nursing staff (42% of all healthcare providers). Depending on their qualifications, these workers belong to three groups: 1) registered nurses (70.6%); 2) licensed practical nurses, (18.8%); and 3) unlicensed personnel (10.6%). As their large number makes them the most costly group of healthcare employees, nurses have become a popular target for cost-containment measures. These include staff cuts, or the replacement of registered nurses with licensed practical nurses, who are themselves replaced by unlicensed aides.
For over three decades, researchers have investigated the complex relationship between nurse staffing, i.e., the quantity of nursing care available, and quality of care received by patients. Appropriate (i.e., safe, sustainable) nurse staffing levels, are associated not only with high quality of care, but also with greater patient satisfaction and lower patient morbidity rates. Nurse managers and administrators still struggle to determine nurse numbers and skill mixes that will safely and reliably achieve optimal patient outcomes. Considering the pressure exerted on those same groups by ongoing cost-containment measures, they urgently need detailed data on this topic.
Part of the reason for the current information shortfall is that research on nurse staffing and its links to patient outcomes has long been constrained by one or more of four limiting factors: 1) aggregated data; 2) cross-sectional design; 3) comparison of high- vs low-performing hospitals; and 4) a lack of relevant details, e.g., shifts, time information, or patient turnover. To our knowledge, to date very few longitudinal studies – and none in Switzerland – have used detailed analyses to examine the association between nurse staffing and patient outcomes. Replication of these studies is needed to confirm both their results and the transferability of their methods.
Overall, this dissertation is structured in six chapters.
Chapter 1 provides an overview of the Swiss healthcare context with associated costs and cost-containment measures. A description and overview of the literature on nurse staffing is presented in terms of nurse education, experience, and skill mix (along with measures to calculate it). An overview of patient outcomes, specifically adverse ones, is then provided concerning incidence, costs, tools for measuring them, as well as the links between some of them and nurse staffing. Finally, short descriptions of nurse outcomes and the Swiss context for nurse staffing are provided. The chapter ends by summarizing the current state of research, including the gaps in the literature, alongside this dissertation‘s contribution to bridging those gaps.
This dissertation’s main aim was to examine the association of nurse staffing with adverse patient outcomes on the shift, unit, and patient levels with a longitudinal design. Chapter 2 gives a detailed description of this aim, which was divided between three studies. In order to achieve this aim, two routine administrative data sources from one Switzerland’s five university hospitals were used for a three-year period (2015 to 2017). The studies’ findings are reported in Chapters 3 to 5.
In Chapter 3, the aim was to describe current study methods and challenges regarding the use of automatic trigger tool-like detection methods via a systematic review. A total of 11 studies met all criteria. The results showed broad variation in applied methods, selection and definition of triggers, estimates of adverse event prevalence, and positive predictive values. Across all 11 studies, adverse event prevalence ranged from 0% to 17.9% (median: 0.8%). The positive predictive value of all triggers to detect adverse events ranged from 0% to 100% across studies, with a median of 40%. No clear evidence was found supporting the development of either a semi- or a fully-automated method of detecting adverse events in electronic health records using trigger tools. I.e., we found no evidence on how to apply or adapt any level of automation to detect adverse events in routine administrative data. For this reason, we chose to use mortality as an adverse patient outcome for Chapter 5.
Before exploring the relationship between nurse staffing and the selected patient outcome (i.e., mortality), a descriptive analysis was conducted with the two merged routine data sources (in Chapter 4). The aim was to longitudinally analyze fluctuations in patient count, nurse count, patient-to-nurse ratio, extreme patient-to-nurse ratio, and patient turnover at 30-minute intervals over a three-year period with a descriptive approach. Our final dataset included more than 58 million data points on 128,484 patients and 4,633 nurses across 70 units. The number of patients showed the highest variability, whereas the number of nurses varied mainly between shifts. Variations occurred between departments, but also within department (i.e., between units), with Intensive Care showing the most stable levels of both nurses and patients. Observing raw numbers of patients and nurses as well as patient-to-nurse ratios and patient turnover figures allowed us to detect the source of that stability. A stable number of patients on a unit does not mean that no patient turnover occurred, but that entries are equal to exits. The same can be said for a stable patient-to-nurse ratio. Additionally, the percentages of shifts subject to extreme staffing situations varied widely across individual departments, quite effectively showing differences in their approaches to maintaining desirable patient-to-nurse ratios. Extreme staffing fluctuations ranged from fewer than 3% (mornings) to 30% (evenings and nights), with percentages falling within “normal” ranges ranging from fewer than 50% to more than 80%. This study provided us with the range of detailed variables on unit-, shift-, and patient-level fluctuations used for the modelling approaches in Chapter 5.
Our final aim, presented in Chapter 5, was to explore the relationship between nurse staffing and patient outcomes. As shown in Chapter 3, as we could not extract any evidence to develop our own semi- or fully-automated adverse event detection methods, we selected mortality as a patient outcome present in our routine administrative data. From our logistic mortality model, we found consistent results for registered nurses: shifts with higher staffing had lower odds of mortality all nights (weekday nights: -17.8%; weekend nights: -13.5%) and weekend mornings (-15.5%), while shifts with lower staffing had higher odds of mortality on weekday nights (+3.9%) and evenings (+30.6%). Findings for licensed practical nurses and Others (including unlicensed and administrative personnel) were not consistent. Unusually high licensed practical nurse staffing was associated with lower odds of mortality for weekday evenings (-5.2%), unusually low staffing with lower mortality nights (-26.4%). For Others, staffing yielded similar results to registered nurses for weekday mornings, where high staffing correlated with lower mortality odds (-3.7%), low staffing with higher odds (+7%). For weekend evenings, though, the association was inverted, with high Other staffing linked to higher mortality odds (+20.3%) and low staffing to lower odds (-6.9%).
To return to the results regarding the registered nurse sample, this high-granularity longitudinal approach supports the hypothesized association between high registered nurse staffing levels and improvement regarding the selected patient outcome. Additionally, the results provide details regarding shifts and other periods during which nurse staffing levels impact patient mortality. As these results eliminate various alternative explanations for the association between nurse staffing and patient outcomes, they also leave a strong likelihood of a causal link.
The sixth and final chapter (Chapter 6) both synthesizes these three individual studies’ major findings and discusses the methodological strengths and limitations of the dissertation as a whole. Moreover, based on the findings and their implications, it includes recommendations for further research, clinical practice and policy. Overall, via its longitudinal design and unit-, shift-, and patient-level data, this dissertation contributes to the understanding of the association between nurse staffing and patient mortality. Its rich granularity attests to the value of the few studies in nurse staffing field to have used this design and methodology. More importantly, it provides insights that apply specifically to nurse staffing in the Swiss context.
The high granularity of our descriptive overview showed variations mainly in numbers of patients – through patient turnover – but also in those of nurses between shifts. Additionally, our exploration of extremes revealed that, since extreme situations occurred quite frequently, even in the apparently well-staffed context of this analysis, many departments struggled to maintain appropriate patient-to-nurse ratios. This was also the case when looking at unusually low- and high-staffed shifts. Unusually low-registered nurse shifts (for weekday nights and all evenings) were associated with an increase in mortality risk, and unusually high registered nurse staffing (for all nights and weekend mornings) with lower risk. This suggests a possible causal relationship between RN staffing levels and mortality.
Based on data drawn from the same sources, licensed practical nurses’ and Others’ contributions to patient safety are nebulous – a result that does not support substitution of either of these groups for registered nurses. Clarification of this and other relationships will require direct examinations – preferably using longitudinal design and shift-, unit-, and patient-level analyses – both of the independent effects of nursing support staffing and of additional adverse patient outcomes more directly linked to nursing care (e.g., pressure ulcers).