Age effect on the prediction of risk of prolonged length hospital stay in older patients visiting the emergency department: results from a large prospective geriatric cohort study

A growing number of older adults (i.e., age 65 and over) visit the emergency departments (EDs) [1]. In Europe, they account for around 20% of all EDs visitors [1, 2]. These older ED visitors, particularly the oldest group (i.e., age 85 and over), generally have a longer length of hospital stay (LHS) after their ED discharge to acute care wards compared to younger ED visitors [1‐3]. The high morbidity burden and related-disabilities expose older patients to an increased risk of non-fatal health outcomes like a long LHS [2, 4, 5]. With the rapid growth of the oldest segment of ED visitors, hospitalization after an ED admission is expected to be even greater in the future and, thus, hospitals need to confront this new challenging issue [1‐3, 6].

One way to reduce LHS is early identification of older ED visitors at greater risk of prolonged LHS after an ED discharge to acute care wards [1, 5, 6]. This screening is a crucial step for targeting appropriate interventions to prevent or decrease the occurrence of non-fatal health outcomes. The predictive tools designed for this purpose should provide a relevant stratification of risk and give information early; ideally before the hospital admission in order to avoid or plan the admission [4, 7]. The use of clinical information collected by a physician has been shown to be the best strategy to develop predictive tools of unplanned hospital admissions compared to self-reported and administrative data collection [8, 9] A limited number of studies have used tools aimed at identifying older patients at greater risk of prolonged LHS after an ED visit, with low predictive accuracy [2, 3, 5, 6]. Recently, the 10-item Brief Geriatric Assessment (BGA), was reported to have a high specificity (97%) but a lower sensitivity 63% [10]. This study reported the best criteria performance to date. This result was explained in part by the use of artificial neural networks (ANNs), and in particular the modified multilayer perceptron (MLP) [10]. Indeed, ANNs analysis is particularly adapted to predict an inherent complex event like prolonged LHS [11, 12]. The main limit of this previous study was the unbalance between sensitivity and specificity, which could be related to the high amount of data required by ANNs [11, 12]. In addition, because the risk of hospitalization increases with age, it could be suggested that the best balance with greater values of performance criteria could be reported specifically in the oldest age group (i.e., age 85 and over) of ED users [1‐4]. The reported study aims to examine the effect of age on the predictive abilities (i.e., sensitivity, specificity, positive predictive value [PPV], negative predictive value [NPV], likelihood ratios [LR], area under receiver operating characteristic curve [AUROC]) of the 10-item BGA for the prolonged LHS using MLP in geriatric ED visitors.

There was a trend for a greater mean age (P = 0.0699), a greater prevalence of temporal disorientation (OR = 2.65, P < 0.001) in participants with prolonged LHS compared to those with short LHS (Table 2). In addition, participants with prolonged LHS visited the ED less often for digestive diseases (OR = 0.48, P = 0.0189) and more often for other diseases (OR = 1.46, P = 0.089) compared to those with short LHS. The mean LHS was 21.6 ± 8.8 days for older ED users with prolonged LHS and 5.2 ± 3.7 days for those who had no prolonged LHS. Whatever the age group considered, predictive performance were high(sensitivity> 89%, specificity≥ 96%, PPV > 87%, NPV > 96%, LR+ > 22; LR- ≤ 0.1 and AUROC> 93). Participants over 75 years showed the best performance (sensitivity = 89.7%, specificity = 97.8%, PPV = 93.4, NPV = 96.5, LR + =41.0; LR- = 0.1 and AUROC = 93.7) (Table 3).

The findings show that effect of age was minimal on predictive abilities of the 10-item BGA. These results suggest that analysis provided by ANNs may enable to use 10-item BGA as a screening tool but also as a predictive tool to identify older patients at higher risk of prolonged LHS, whatever their age.

The best criteria performances for prolonged LHS were shown with patients aged 75 years and over. This is an unexpected finding because it was hypothesized that greater values of criteria performance could be reported in the oldest segment of ED users. This result is discordant with previous studies which reported a strong association between age and the risk of prolonged LHS [1‐4]. Age has previously been identified as an important predictor for prolonged LHS [10, 13, 14]. For instance, in a similar sized cohort of patients admitted to ED (993 patients, mean age = 87.04 years) age and gender explained 21.6% of area under receiver operating characteristic curve value [10]. In the same way, Campbell et al. reported in a larger cohort of patients admitted to ED (1626 patients, mean age = 78.7 years) that age over 85 years was strongly associated with prolonged LHS (OR = 7.6, P < 0.001) [14]. The association between increased age and prolonged LHS has been explained by incident disabilities that exceed 50% in hospitalized patients aged 85 years and over [14‐16]. This finding is consistent with Sourial et al. who showed that age and gender had the highest contribution (C statistic values from 0.51 to 0.67) in predictive accuracy of incident disability in a cohort composed of 6657 patients (mean age = 73.68 years) [17].

A possible explanation of the discordance about age effect on the prediction of prolonged LHS shown in our study compared to previous studies could be related to the profile of population recruited, which is oldest old patients with a mean age around 85 years. Moreover, ANNs provide a different statistical approach that consider the complex interplay between all items [11, 12, 18]. Indeed, previous results of ANNs reported that using numerous variables increased predictive accuracy (area under cover values lie between 84.1 with 9 items and 90.5 with 10 items) but also modified the contribution of demographic items in the predictive performance (from 12.8% with 9 items to 21.3% with 10 items) [10]. Categorizing age groups provide an additional variable that limits the analysis of ANNs to a single age group and may modify the distribution and weight of 10 items in contribution of the predictive accuracy. Thus, ANNs take into account the variations in the contribution of all types of variables (demographic items, acute or chronic diseases, and environmental items) to increase predictive performance and to learn to recognize patterns of prolonged LHS in each age group.

Our findings underscored that regardless of age, values of criteria performance were high (sensitivity> 88%, specificity≥ 96%, PPV > 87%, NPV > 96%, LR+ > 22; LR- ≤ 0.1 and AUROC> 93). To the best of our knowledge, the current study demonstrates the best performance and balance between criteria reported for prediction of LHS after an ED visit. This result is discordant from a recent previous study which reported lower values and an unbalance between sensitivity and specificity [10]. The main explanation as suggested in our hypothesis could be related to a difference in the number of participants. In our study, we included 1117 individuals which likely increased the accuracy of prediction. It has been shown that ANNs may provide accurate information on an event only if there is a sufficient quantity of data points to be analysed [11, 12].

In order for a screening test to be applicable, it requires high level sensitivity to limit false-negative results. With sensitivity above 89% in all age group conditions, the 10-items BGA is a useful screening tool that can be applied to identify early older ED users at higher risk of prolonged LHS after their discharge to acute care wards. In addition, our results demonstrated a high specificity above 95%, which implies that there is a low false positive rate when applying the 10-items BGA to predict length of hospitalization. Thus, the combination of both high sensitivity and specificity indicates that the 10-items BGA is not only a simple screening tool but also a diagnostic tool with excellent predictive accuracy. The ability to analyse our data with the use of ANNs methods analysis is the main explanation for these findings. Indeed, ANNs are data analysis tools developed to overcome the limitations of traditional linear models as a method to predict health events [18]. ANNs are computational models which are capable of machine learning and pattern recognition [7‐12]. Because they apply non-linear statistics to pattern recognition, ANNs are particularly adapted to “chaotic” events like prolonged LHS. Nowadays, the advances generated by ANNs combined with improvement of computer technology affords the opportunity to explore new perspectives using ANNs as decision-making diagnostic aids for physicians. Thus, these results can be applied directly to clinical practice because they can be developed as software applications for computers and hand held devices. The 10-item BGA may provide answers to facilitate clinical decision-making process because this tool provides a risk stratification of patients at risk of non-fatal health outcomes. Such information may be relevant to make the right decision for the patients like the discharge to home or to a medical ward, and to continue the appropriate interventions in the right patients and at the right time by the right professionals (i.e., geriatric intervention versus no geriatric intervention).

Our results also showed that patients with digestive diseases had a shorter LHS compared to patients admitted for other diseases. One explanation could be that admissions for digestive diseases are more often semi-urgent or non-urgent [19]. This low degree of urgency may explain a lower LHS. In contrast, “Other diseases” as a reason for admission was associated with an increases LHS. This group refers in part to traumatic injuries related to falls. Unlike young patients, the most common mechanism for traumatic injuries in older patients is due to fall [20]. Falls have been identified as major cause of unintentional injury leading to prolonged LHS and death, especially over 80 years, that could explain our results [21].

The strengths of this study include the large number of participants, the prospective cohort design, the hard outcome represented by prolonged LHS, and the use of sophisticated new statistical models. However, limitations need to be considered including recruitment of participants from a single center and the fact that important items related to prolonged LHS could have been forgotten. Besides, we included inpatients who died during their hospitalization and those discharged in another hospital. Both date of death or transfer to another hospital were considered as the last day of hospitalization. Thus, a bias might exist because of a suspected higher complexity of those patients.

Age stratification provided minimal effect on the abilities of the 10-item BGA to predict prolonged LHS using ANNs. Indeed, ANNs provided homogeneous predictive performance that enable to use 10-item BGA as a screening tool but also as a predictive tool to identify older patients at higher risk of prolonged LHS, whatever their age.