A Simple and Accurate Model for Predicting Fall Injuries in Hospitalized Patients: Insights from a Retrospective Observational Study in Japan

Background

Fall injuries during hospitalization increase the cost of health care [1–4] due to the increased cost of hospitalization caused by extended hospital stays and the accompanying increased cost of treatment [1,5–6]. In addition, legal resolutions of falls considerably and seriously burden health care providers [7]. However, among the patients who fall while in hospital, 88% suffer no injuries and do not require any additional treatment [8]. Meanwhile, patients who do incur fall injuries are susceptible to limited activities of daily living (ADLs) or other long-term sequelae [9,10]. Therefore, it is more important to predict the occurrence of fall-related injuries than a fall itself without concomitant injuries. While many medical institutions in Japan use various tools to prevent a fall or fall injuries in inpatients, such as bed alarms, low beds, or hip protectors, medical resources available for such tools are limited [11–13]. Thus, the target population to be offered such preventive tools must inevitably be narrowed down to those patients more likely to have fall injuries. However, given that a reliable model to predict such patients remains to be established, medical practitioners in many medical institutions select their targets for intervention according to their previously developed risk assessment tools for falls. Therefore, development of a model that can accurately predict which patients are likely to have fall injuries would help medical practitioners worldwide to allocate finite human and material resources more effectively and efficiently.

Several predictive models for falls regardless of concomitant injuries have been reported, such as the St. Thomas Risk Assessment Tool, the Tinetti Mobility Test, the Hendrich II Fall Risk Model [14–16], and the Saga Fall Prediction Model we reported in 2020 [17]. In addition, studies targeting only patients who fall have reported various factors associated with fall injuries, such as being elderly, female, or visually impaired, and scoring rank A of bedriddenness ranks (BRs) from official ADLs scales of the Ministry of Health, Labour and Welfare (MHLW), Japan [8,18–20]. However, there have been no published reports on factors or models predicting fall injuries targeted to all adult inpatients. Because the risk of falling is reported to be highest within 1 week of admission [21], the predicting factors constituting a predictive model for fall injuries should be easily available on admission. Therefore, this retrospective study aimed to develop a predictive model for fall injuries targeting all inpatients aged ≥20 years by using simple predictors that were easily assessed on the day of admission.

Material and Methods

ETHICS STATEMENT:

This study complies with the Ethical Guidelines for Medical and Health Research Involving Human Subjects of the MHLW and the Ministry of Education, Culture, Sports, Science and Technology, Japan. This study was approved by the Research Ethics Committee of Yuai-Kai Foundation and Oda Hospital (No. 20180629_R1). Consent of all patients was obtained by the comprehensive consent method in the hospital, and patient anonymity was protected.

STUDY DESIGN AND POPULATION:

This was a single-center, retrospective cohort study. We enrolled all inpatients aged ≥20 years admitted to an acute care hospital (Yuai-Kai Foundation and Oda Hospital) from April 2012 to March 2018. The hospital is located in the rural city of Kashima, Saga Prefecture, Japan, serving a population of about 90 000. More than 3100 inpatients are treated annually in the hospital, with average length of stay of 12.1 days. The hospital comprises 10 departments and 111 beds, including the Departments of Internal Medicine, Surgery, Neurosurgery, and Cardiovascular Surgery.

VARIABLES AND SOURCES OF INFORMATION:

The data in this study were extracted from the hospital’s health records. The variables extracted have been reported as factors related to a fall or fall injuries in previous studies: age (continuous variable), sex (male or female), department of admission (Internal Medicine, Neurosurgery, or others), emergency admission (presence or absence), emergency transport (presence or absence), medical referral letter (presence or absence), history of falls (presence or absence), use of hypnotic medication (using or not using), visual impairment (presence or absence), parkinsonism (presence or absence), permanent damage by stroke (presence or absence), ability to eat, swallow, toileting, hold sitting, hold standing, transfer from bed to wheelchair/stretcher/portable toilet, move, bathe, and take prescribed medicine (independent or requiring assistance), and MHLW classifications for ADLs; BRs (normal, J, A, B, C), cognitive function scores (CFs) (normal, 1, 2, 3, 4, M) on admission and rehabilitation (presence or absence), surgery (presence or absence), fall (presence or absence), injury level caused by a fall (no injury and levels 1, 2, 3) [22], and dates of admission to and discharge from hospital. Among the variables we extracted, 8 variables – age, sex, emergency admission, department of admission, use of hypnotic medication, history of falls, ability to eat, BRs – had been reported as indicators in a predictive model for fall regardless of having concomitant injuries [17]. A sample size could not be calculated because a predictive model for fall injuries had not been previously reported.

DEFINITION OF FALL AND FALL INJURY:

The case of a patient’s fall during hospitalization is usually recorded in an incident/accident report by the attending nurse or the nurse who discovered the fall. “Falls” were defined as unexpected falls from any unplanned descent of a patient to the floor with or without injuries, from any height, and in any position, ie, on the floor, from stairs, chair, or bed, when standing, walking, or sitting, or even in recumbent position. Whether the patient had a fall or not was identified from the incident/accident reports described above. The severity of an injury caused by the fall was classified according to a part of “Reasons for Falls and Falls and Injury Severity Input Criteria” in the acute care indicators in the Maryland Hospital Association Center for Performance Sciences’ Quality Indicator Project (Table 1) [23]. We divided inpatients into 2 groups: the group “With fall injury” involving the patients with severe injuries more than injury level 1, and the group “Without fall injury” involving the patients who had not fallen or had fallen without having any injuries.

DEFINITION OF VARIABLES:

The variables history of falls, use of hypnotic medications, visual impairment, parkinsonism, permanent damage by stroke, ADLs, BRs, and CFs were routinely assessed by the attending nurse within 24 h after admission. MHLW BRs and CFs are official tools for assessing ADLs that have been used widely under the long-term care insurance system in Japan. BRs are classified into 5 major categories (normal; J, independence; A, house-bound; B, chair-bound; C, bed-bound) (Figure 1). CFs are classified into 6 major categories (normal; 1, independence; 2, attention needed; 3, intermittent care needed; 4, continuous care needed; M, specialized treatment needed) [8,22]. In the hospital, hypnotic medications were classified into benzodiazepines and nonbenzodiazepines, excluding melatonin receptor agonists and orexin receptor antagonists. Although “visual impairment” had not been defined clearly in this retrospective study, attending nurses determined a patient as being visually impaired when the patient’s ADLs were limited by some form of visual abnormality.

STATISTICAL ANALYSIS AND METHODS OF CREATING THE PREDICTIVE MODEL:

Univariable analysis for fall injury was performed by a chi-squared test with calculated odds ratio (OR), 95% confidence interval (95% CI), and P value. Continuous and categorical variables are presented as a median (interquartile range) and an absolute number (percentage). The factors that were available on admission were selected as candidate covariates in multivariable regression. In terms of collinearity, we selected 1 factor among the candidate variables that showed a high correlation coefficient (Spearman’s r >0.7) with the others as a covariable in the multivariable model, then excluding the others. The predictive model was created using logistic regression analysis (Analysis 2) by forced entry method using the covariates that were significant in another logistic regression analysis (Analysis 1) by forced entry method before Analysis 2. In the internal data set, the discrimination of the model was assessed by the area under the curve (AUC), and the calibration of the model was assessed by shrinkage coefficient. Both sensitivity and specificity were calculated at the 3 cutoff values of the model, with sensitivity of 90%, specificity of 90%, or maximum sum of sensitivity and specificity (Youden’s Index). Analyses were performed using SPSS version 25 (SPSS, Chicago, IL, USA).

Results

CHARACTERISTICS OF THE PATIENTS AND INCIDENCE OF FALL INJURY:

During the study period, 17 687 inpatients were admitted. A total of 17 062 patients were analyzed, excluding 624 patients <20 years old and 1 patient without a record of outcome of fall injuries during hospitalization. The median age (interquartile range) was 77 (63–85) years; 8399 (49%) patients were men, and the median length of hospital stay (interquartile range) was 9 (4–17) days. There were 646 (3.8%) patients who had a fall during hospitalization (the incidence rate of falls was 5.33 per 1000 patient-days), and 113 patients were classified into the group “With fall injury” (0.7% of all hospitalized patients, or 1.21 per 100 000 patient-days, Figure 2). In the “With fall injury” group, the median age (interquartile range) was 85 (80–89) years, and 64 (57%) patients were male.

UNIVARIABLE ANALYSIS:

The results of univariable analysis are shown in Table 2. Patients with fall injuries were significantly older (85 vs 77 years) and had longer hospitalization (22 vs 9 days). In addition, the rates of the following factors were significantly higher in the “With fall injury” group: admission to Internal Medicine (64% vs 46%), emergency admission (69% vs 45%), absence of emergency transport (67% vs 40%), presence of medical referral letter (43% vs 29%), history of falls (30% vs 10%), requiring assistance in eating (35% vs 21%), requiring assistance in toileting (46% vs 27%), independence of sitting (73% vs 61%), requiring assistance in standing (30% vs 24%), independence of transfers from bed to wheelchair/stretcher/portable toilet (62% vs 42%), requiring assistance in moving (49% vs 31%), requiring assistance in bathing (59% vs 33%), BRs of J (12% vs 9.1%), A (38% vs 15%), and B (25% vs 11%), CFs of 1 (21% vs 10%), 2 (9.7% vs 3.1%), and 3 (35% vs 15%), undergoing rehabilitation during hospitalization (67% vs 40%), and without undergoing surgery during hospitalization (79% vs 69%).

COVARIATE SELECTION AND MULTIVARIABLE ANALYSIS:

Among all of the 24 collected factors, except for 3 indicators (surgery, rehabilitation, and length of hospital stay) that were impossible to access at the time of admission, the following 13 factors were selected as covariates for multivariable analysis, taking their collinearity into consideration: age (continuous variable), sex (male or female), department of admission (Internal Medicine, Neurosurgery, or others), emergency admission (presence or absence), emergency transport (presence or absence), medical referral letter (presence or absence), history of falls (presence or absence), use of hypnotic medications (using or not using), visual impairment (presence or absence), parkinsonism (presence or absence), permanent damage by stroke (presence or absence), ability to eat (independent or requiring assistance), and BRs (normal, J, A, B, C). The results of the binomial logistic regression analysis (forced entry method, Analysis 1) showed that 6 indicators were significantly associated with fall injuries (Table 3, Analysis 1): older age (OR 1.0, 95% CI 1.0–1.1, P=0.003), male (OR of female 0.5, 95% CI 0.4–0.8, P=0.001), absence of emergency transport (OR of presence 0.4, 95% CI 0.2–0.7, P=0.003), presence of medical referral letter (OR of presence 1.5, 95% CI 1.0–2.2, P=0.042), presence of history of falls (OR of presence 1.7, 95% CI 1.1–2.7, P=0.013), abnormal BRs: BR J (OR of J to normal 4.7, 95% CI 1.9–12, P=0.001), A (OR of A to normal 7.9, 95% CI 3.5–18, P<0.001), B (OR of B to normal 6.2, 95% CI 2.6–15, P<0.001), and C (OR of C to normal 4.1, 95% CI 1.6–11, P=0.004). Similar to Analysis 1, in the binomial logistic regression analysis (forced entry method, Analysis 2) using only the 6 covariates that were significant in Analysis 1, all of the 6 factors were significantly associated with fall injuries (Table 3, Analysis 2): older age (OR 1.0, 95% CI 1.0–1.1, P=0.001), male (OR of female 0.5, 95% CI 0.4–0.8, P=0.001), absence of emergency transport (OR of presence to absence 0.4, 95% CI 0.3–0.6, P<0.001), presence of medical referral letter (OR 1.5, 95% CI 1.0–2.2, P=0.041), presence of history of falls (OR 1.7, 95% CI 1.1–2.6, P=0.012), abnormal BRs: BR J (OR of J to normal 4.9, 95% CI 2.0–12, P=0.001), A (OR of A to normal 8.9, 95% CI 4.0–20, P<0.001), B (OR of B to normal 7.3, 95% CI 3.1–17, P<0.001), and C (OR of C to normal 4.5, 95% CI 1.9–11, P=0.001).

The predictive model for fall injuries we developed using the regression coefficients in Analysis 2 is as follows:

Score=−8.672+0.034 × (age) + (sex: female=−0.641) + (emergency transport: presence=−0.887, missing=−0.551) + (medical referral letter: presence=0.399, missing=1.189) + (history of falls: presence=0.542) + (BR: J=1.597, A=2.188, B=1.986, C=1.510, others=0.154).

INTERNAL VALIDATION OF PREDICTIVE MODEL:

The AUC as the predictive performance of this model calculated using data from the same population was 0.794 (95% CI 0.762–0.826, Figure 3). The cutoff values for the model with the 90% sensitivity and 54% specificity points was −5.27, and with the 35% sensitivity and 90% specificity points was −3.80. The score of the cutoff value with Youden’s Index was −5.08, with sensitivity 89% and specificity 56%. Predicted and measured incidence rates of fall injuries in each of 10 quartiles of the score are shown in Figure 4. The shrinkage coefficient was 0.955.

Discussion

We developed a new model to simply predict fall injuries during hospitalization using 6 predictors that were routinely obtained on admission in acute care settings in Japan. As predictive models for fall injuries have yet to be reported, the model developed in this study is the first predictive model for fall injuries to target all adult inpatients. Fall injuries during hospitalization lead to increased health care costs [1–4] and additional limitations of patients’ ADLs with possible long-term sequelae [9,10]. Although using hip protectors or multifaceted interventions based on risk assessment for a fall was reported to be effective in preventing a fall or fall injuries [13,24], it is unrealistic to apply such resources to all inpatients who are at risk of a fall in the setting of actual acute medical practice, owing to the limited supply of such equipment, medical resources, and time required. Therefore, several predictive models for a fall during hospitalization were developed and are duly employed, but without taking accompanying injuries into consideration [14–17]. However, 88% of patients who fell during hospitalization were reported to be uninjured without needing any additional treatment [8]. This fact suggests that preventive intervention should make more efficient and appropriate allocation of finite human and material resources to patients with a high probability of an injurious fall rather than to those at risk of a fall without concomitant injuries. In addition, assessment of the risk of fall injuries as early as possible is preferable because the highest incidence of falls is reported early after admission, especially within 1 week [21]. Considering these points, our prediction model focusing on the prediction of fall injuries rather than a fall itself regardless of the occurrence of concomitant injuries, which could be assessed immediately after admission using 6 simple predictors (age, sex, emergency transport, medical referral letter, history of falls, and BRs) available on admission, could be extremely useful in preventing fall injuries among inpatients in acute medical settings.

We used BRs, one of the official Japanese ADLs classifications, as one of our predictors. BRs have been widely used as a scale of disability among the aged population in medical and long-term care settings in Japan, and have also been used in several studies [8,25,26]. BRs classify the degree of ADLs decline roughly into 5 ranks through 3 steps [22], which is much easier to assess than the Katz Index with 7 ranks of overall rating requiring assessments of 7 items with 2 choices [27], or the Barthel Index requiring assessments of 10 items with 2 to 4 ranks [28]. BRs were also recently reported to have good inter-rater reliability and criterion-related validity, similar to the Katz Index and Barthel Index [22]. Thus, BRs are reasonable predictors to use in a predictive model that requires simplicity of use in clinical practice.

The multivariable analysis of this study identified the 6 variables – age, male, absence of emergency transport, presence of medical referral letter, presence of history of falls, and abnormal BR J, A, B, or C – as significant factors associated with fall injuries. The fact that age and BR of A are useful predictors for fall injuries is consistent with results previously reported [8,19,20], and in another study male was inconsistent [18]. Other remaining factors – absence of emergency transport, presence of medical referral letter, presence of history of falls, and abnormal BR J, B, or C – were new predictive factors for fall injury. The study that reported female as a predictive factor for fall injury was conducted only for patients who had fallen [18], which was different from our study that included patients who had not fallen. Male [16,17,29,30], presence of history of falls [31], and BR B or C had also been reported as factors associated with a fall itself regardless of having concomitant injuries [17], suggesting that these 3 factors could have predicted a fall itself, which was a matter-of-course prerequisite for having injuries. In addition, BRs are useful for predicting fall injuries among inpatients because of their higher OR for predicting fall injuries in comparison with other indicators in the model.

This study also showed that the absence of emergency transport and presence of a medical referral letter were 2 novel factors associated with fall injuries. The reason that the absence of emergency transport was associated with fall injuries could be the lower incidence of using physical restraint in such situations, which would naturally prevent a fall itself and subsequent related injuries. Physical restraint was reported to be more often used for young patients with psychiatric disorders or alcohol dependence or, on the contrary, aged patients with internal diseases, both of which tended to be transferred by ambulance [32]. Additionally, physical movements of patients being transferred by ambulance are often restricted because of their medical conditions or treatment modalities, such as bed rest after surgery or use of ventilator or infusion pump [33], which possibly lowers the incidence of a fall and fall injuries. In fact, the association between transported by ambulance and severity of fall injury was not significant in our previous study [8]. However, the proportion of patients transported by ambulance was markedly lower in the group with severe fall injuries than having mild or no injuries among inpatients who fell, which could mean that an insufficient sample size caused lower statistical power, thus preventing a possible significant difference. Concerning the other new factor, the presence of a medical referral letter, it is possible that it might have predicted only a fall without regarding concomitantly occurring injuries, consequently showing its significant relationship with fall injuries in this model. The percentage of patients having a medical referral letter with a fall in our previous study was higher (37–39%) than that of patients without a fall in another previous study we reported (33%) or in this study (29%) [8,17], although in our previous report the patients who fell showed no significant difference in the ratio of the presence of a medical referral between the patients with and without injuries [8]. Further research is required to decide whether the presence of a medical referral letter is an accurate predictor of a fall or to elucidate the reasons for the fact that a medical referral letter is associated with a fall, because all our studies already mentioned were conducted at the same institution in Japan, which was an unavoidable limitation.

This study has several limitations. There could be an overestimation bias because this study was a single-center, retrospective study without external validation, owing to the small sample size of patients with fall injuries. External validation using data from different populations with various backgrounds will be necessary to determine the proper prediction power and generalizability of this model because some predictors, including presence of emergency transport or medical referral letter, could have diverse indications in different medical settings, such as at different hospitals with various regional medical backgrounds. In addition, the ORs for fall injuries could have been underestimated because the influence of fall prevention procedures was not adjusted. These had been routinely implemented in the hospital according to its original risk assessment tool due to the failure of data collection.

Conclusions

We developed a new prediction model for fall injuries among inpatients using 6 predictors easily available on admission to hospital, including the subcategory BRs of the official ADLs scales in Japan. This is the first predictive model for fall injuries during hospitalization of adults in general hospitals, which could be useful for selecting priority targets of patients for fall prevention in busy clinical settings because of its simple and quick risk assessment.