This research was a secondary analysis of data collected for a study that considered the costs and consequences of different nursing staff configurations (Trial Registration: ClinicalTrials.gov NCT04374812) [9]. That study had a waiver for informed consent after assessment via the Health Research Authority integrated research application service (IRAS ref 273185). The ethical review undertaken by the faculty research ethics committee at the University of Southampton (ERGO 52957) for that study was extended for this secondary analysis on the basis that its objectives were aligned with the purpose for which the data had been obtained. The dataset had previously been anonymized with no linkages back to individuals so privacy and confidentiality were respected. The hospital trusts that provided the data gave approvals for this secondary analysis. Dataset construction and all analyses were performed using R (version 4.4.0; R Core Team); further details of the software and packages used can be found in Multimedia Appendix 1.

We analyzed data from adult inpatient wards in 1 hospital trust for admissions from February 2017 to February 2020. The data consisted of anonymized admission records and patient classification assessments made using the SNCT acuity instrument. SNCT acuity assessments had been undertaken in admission units, general wards, and high-care wards but not in intensive care units. The admission units in this hospital were the medical emergency and surgical emergency assessment units, which assess patients before transferring them to a general or high-care ward as required. In the UK, a general ward is an inpatient adult ward not specializing in maternity or psychiatric care. A high-care ward has some high-care beds which are defined as providing level 2 care according to the UK intensive care society, deemed as needing 0.5 WTE of nursing care per bed. A UK intensive care unit, on the other hand, provides the highest level of care, level 3, and this requires one-to-one nursing care. The SNCT acuity assessments data provided counts of patients in a ward for each acuity and dependency category at a point in time. The admission records included each patient’s first NEWS (National Early Warning Score) [10]. NEWS is a patient assessment method in routine use in UK hospitals based on vital signs observations, which nurses use regularly to identify patients at risk of deterioration.

The patient categorization instrument used for ward workload assessments in SNCT involves matching patients to 1 of 5 ordinal categories according to descriptions of acuity and dependency needs. Level 0 is for stable patients with needs met in general wards, level 1a patients are acutely ill patients requiring intervention or those with a greater potential to deteriorate, level 1b patients are stable but are more dependent on nursing care, and level 2 patients are unstable, at risk of deteriorating, and require specialized care. Level 3 patients, who require multiple organ support, did not exist in our dataset of assessments. Each category has a value (called a multiplier) corresponding to a staffing WTE needed to provide 24-hour care after considering annual leave and sickness absence. Multipliers were developed from ward observational studies [11]. One set of multipliers is used for general wards and another for admission units to take into account the extra workload associated with the higher throughput of patients. The multiplier values for the patients in a ward are summed to produce a point-in-time WTE estimate for the ward [4].

All admission records that could be linked by ward and by assessment date and time were included for wards that were routinely carrying out the assessments. Since the assessment records did not record any patient identifiers, linkage to patient records was by ward and time. Linked records were analyzed where the number of patients in the ward assessment matched the number of occupants in the ward according to the patient administration system at the time of the assessment. A comparison of linked and unlinked assessments was conducted to check for bias resulting from this selection process.

The dependent variable for the prediction model was the WTE calculated by the assessment tool divided by the number of assessed patients. Patients’ data for the assessed patients, as determined by the linkage described earlier, were aggregated to ward level by taking means for continuous variables and proportions for binary variables, as shown in the table. The choice of candidate predictors (Table 1) was made according to data availability, published nurse workload models, and clinical protocols regarding nurse assessments for new ward arrivals. Multi-variable models using all the candidate predictors were built since there was evidence or conceptual reasons for including them, and the objective of the modeling was predictive accuracy rather than quantifying effects of individual predictors. Univariable regressions were performed on the candidate predictors out of interest. The predictors include the use of 2 derived variables: the Charlson comorbidity index [12] and the summary hospital mortality indicator (SHMI) risk [13]. The diagnostic predictors were proportions of patients with a recorded diagnosis belonging to a particular diagnostic group: respiratory conditions, cardiac conditions, renal conditions, cancer, frailty syndromes [14], a self-harm diagnosis, and a history of poisoning. The details of the diagnostic groups are provided in Multimedia Appendix 1. Lastly, an indicator was added for admission units to reflect the alternate set of patient category multipliers used for assessments in them.

A generalized additive regression model (GAM) [15] with an identity link function was chosen for the modelling since this could handle nonlinear relationships between the individual predictors and the WTE per patient response variable.

The data were randomly split into training and test sets (75% vs 25%), and a regression model was created on the training set. All the predetermined candidate predictors were used after checking for high pairwise correlations. Removal of any predictor was found to worsen model fit judging by the Akaike information criterion, so all were retained. A calibration was computed to optimize the predictions by means of a smoothing of the model estimates on the assessed ones [16]. The model and the calibration were applied to the test set.

A Bland-Altman plot [17] was constructed and used to look at the agreement between the model prediction and the assessed WTE per patient. When examining this plot by ward subset, 2 wards had more than a third of their predictions outside the limits of agreement. This was a markedly higher proportion of assessments than for the other wards, so these wards were excluded from the datasets, and the GAM model was rebuilt.

In this study, we used a limited set of routine patient data to estimate the demand for nursing staff on general and high-care hospital wards and admission units. We found that data on patient demographics, routes of admission, ward transfers, diagnoses, and admission NEWS, which are available in hospital systems, could be used to estimate the results of assessments made using a PCS with a moderate degree of precision. The predictions of WTE per patient were on average within 5% of the recorded estimates when applied to a test set. Calibration “in the large” was good in the test set, with the mean prediction being only 0.002 WTE per patient less than the recorded mean [18]. However, a calibration curve showed that the model tended to underestimate at higher workloads. This would be the case if some dimensions of patients’ care needs in more acute wards were not being adequately represented in the dataset. A “by ward” sub-analysis of the model estimates showed that the surgical high care unit and the renal transplant unit were the 2 wards where the model underestimated the most (as proportions of predictions). This might indicate the model is lacking variables that adequately capture the acuity and dependency of patients after major surgery. If the model were to be used as it is, a professional judgment framework, such as that published by Saville et al [19], should be used, which might recommend considering adjustment of the estimates where they exceeded 1.75 WTE or where there were high-care beds in the ward.

The predictors were not chosen to be orthogonal; they clearly have correlations when considering their potential effect on nurse workload. That means Figure 2 explaining their contribution to estimating workload needs to be interpreted with care. It is not surprising that the admission unit indicator makes the largest contribution since an approximately 10% uplift in nurse staffing for admission units is baked into the PCS, which determines the target WTE per patient. It is perhaps expected that the age, comorbidity index, and admission NEWS case mix variables of ward occupancy influence estimates of workload. However, the marked contribution of gender invites comment. It may reflect staffing in single-sex wards dealing with gynecological or urological conditions.

While overall fit was good, the Bland-Altman plot shown in Figure 4 revealed a distinct artifact of 45-degree downward-sloping lines, which arises from ranges of continuous prediction values estimating a given workload derived from the discrete set of multiplier values in the PCS and the numbers of patients in wards. There is a degree of subjectivity in categorizing patients in the PCS into one of the 4 levels, and there is a possibility that there was some up-coding of assessments in which patients were put into a higher category, for example, level 1b rather than level zero, which might boost a ward’s chances of getting more resources [20]. The high proportion of patient assessments in the level 1b category shown in Table 2 possibly suggests some of these were erroneous, although patients with dependent care needs may have longer than average stays with opportunity for more categorizations at that level.

In conversations, safe staffing managers and leads at a number of hospitals have expressed concerns regarding the accuracy and reliability of manual patient categorizations by ward staff. An automated workload prediction system would be objective and eliminate or reduce a source of bias. It would also remove the need for training staff in carrying out the assessments and retraining them when assessment methods are revised.

The OPCq and Perroca patient classification instruments for assessing patient care needs are more structured than the SNCT instrument, as they require scoring a patient across several domains before determining an overall category. The drawback to this rigor is that the assessments take longer. One nurse quoted in the publication by Ayan et al (2024) [21] concerning the Perroca instrument said: “Since we have been using it continuously, it does not take me much time now. It only takes five minutes to assess a patient; it couldn’t be easier.”

Our observational data (unpublished) from another study [22] suggests that assessing all patients on a 30-bedded ward 3 times per day with SNCT categorizations could occupy a senior nurse for up to 30 minutes per day, plus extra time for data entry.

The identification of 2 wards with outlier predictions is of interest. In the private patient unit, the model tended to predict higher staffing requirements than those identified in the assessments. This could arise because of a selection effect, whereby the occurrence of more elective procedures on more medically fit patients was not reflected in the variables used for this population. In contrast, the predicted workload estimates for the surgical high care unit tended to be too low, and this might mean the model is not adjusting adequately for these acutely ill patients. Additional predictors to adjust for operational severity might improve the predictions.

The surprising accuracy of the predictions on limited data opens up the possibility of automated near-time assessments and consequently freeing up some nursing time. The feasibility of implementing the prediction algorithm in real-time or near-time is dependent on the data items being recorded and accessible in a timely manner. In the United Kingdom, basic patient admission data is usually recorded in near-time in a patient administration system with electronic interfacing, though the clinical coding of diagnoses in the same system is often much later. However, the diagnoses the algorithm uses are largely for long-term conditions or conditions known on presentation, in which case they may already be present in the electronic clinical history or recorded in nursing or clinical handover documentation. A National Health Service (NHS) England-sponsored digital maturity assessment of UK hospital trusts reports that 32 of 132 acute trusts have electronic nursing and physician documentation [23], and that number is rising, so real-time electronic provision of patient conditions is becoming more widespread.

The approach taken in this study differs from previous studies that built models or made estimates of actual historic staffing levels [24-26]. In contrast, our approach was to build a model of estimated staffing requirements trained on the records of an existing staffing estimation tool [27]. Unlike the classifier model built by Rosa [7], which was trained on values from the PPCI to make patient categorizations, ours was a prediction model of ward staffing requirements. While the prediction of PPCI values was only moderately accurate (72% correct classification), in our study, the magnitude of errors for estimating ward staffing requirements was, on average, small, with a mean average percentage error of less than 5% and 95% of estimates within 14% of the assessed value. The required precision is an open question, but staffing within ±15% of measured demand is used as a criterion to define acceptable variation in one widely used system [28].

The method we have used to build a model that is trained on the ward workload is flexible and would work for other patient acuity and dependency scoring instruments that assign an ordinal value for the patient’s care needs.

The dataset was limited in terms of longitudinal clinical data regarding patient conditions. While it contained the NEWS on admission, it did not have subsequent scores or individual vital signs. Without time-varying variables such as NEWS and other assessments that reflect a patient’s evolving acuity, the model is relying on the ward case mix not varying from the average in terms of proportions of patients medically fit for discharge or proportions deteriorating. These 2 cases work in opposite directions in terms of workload, those patients who are medically fit probably requiring less nursing care, while those who deteriorate probably requiring more nursing care. A systematic bias in the model performance is not easy to determine without further information. Also, as noted earlier in the main findings, the poor performance of the model in the private patient unit and in surgical high care might result from the model lacking information on surgical procedures and operation severity. This limits the settings in which the current model might be used.

Another limitation that might have impacted the accuracy of the predictions was the reliability of the recorded patient classifications. The use of the acuity categorization tool for estimating staffing requirements in the dataset was new to the hospital at the time; hence, the reliability of the assessment may have depended for a period on the effectiveness of training given to ward assessors and monitoring of the use of the tool.

The robustness and generalizability of the model need to be checked by assessing its performance using data from other hospitals. It is possible that variation in patient case mix and in how hospitals apply SNCT patient acuity categorizations could affect model performance. It should also be noted that while the aim of this study was to estimate workload that matched the existing SNCT-based assessments, the tool in use did not refine workload requirements by skill mix, for example, advising how many registered nurses and how many nurse support staff might meet the requirements, nor about requirements for specialist skills. This would be a useful step in matching requirements to available resources and might be the subject of further work.

This research demonstrated a way of building a moderately accurate prediction model of nurse staffing requirements from a limited dataset of routine data by training the model on records of an established patient acuity and dependency patient classification tool. The results demonstrate the potential for automating assessments of nurse staffing requirements from routine data, reducing time spent on this nonclinical overhead, and improving monitoring of real-time staffing pressures.