Observed-predicted length of stay for an acute psychiatric department, as an indicator of inpatient care inefficiencies. Retrospective case-series study.

Today there is a growing interest in improving quality and efficiency of health care to the maximum, a fact which highlights the necessity of good indicators of quality and efficiency of health care.

Length of stay (LOS) has been repeatedly used as an indicator of efficiency for inpatient care, probably due to its clear meaning as one of the main sources of hospital costs and because LOS can be also deemed an indicator of quality [1‐3]. However, LOS for a certain period and facility, is not a useful basis for meaningful comparisons unless it is adjusted for the case mix of patients hospitalized during the period considered. This process is called "risk adjustment" and is thoroughly described and discussed in the book edited by Iezzoni [4].

Allegedly the ideal way of adjustment should be based on the difference between LOS a patient should require–provided the attention received was efficient–, predicted LOS, and the actual one, observed LOS. As a continuous variable, LOS (or a proper transformation) may be modeled by means of a linear regression approach [5]. If an adequate model is found, the difference between observed LOS (OLOS) and predicted LOS (PLOS) could be a proper efficiency indicator. We already evaluated such an indicator for Internal Medicine and Surgery departments with fair results [6].

As a matter of fact, several studies, some of them recently performed, have looked for LOS predictors in psychiatric departments with similar aims [7‐19].

Diagnosis, severity of illness, age, sex, physical co morbidities, treatment issues and psychosocial characteristics have been already confirmed as LOS determinants with more or less strength across the above referred studies.

The study is basically a retrospective case-series study. It is mainly descriptive although some hypothesis testing has been performed during the derivation of the function.

Our results focus on the plausibility of obtaining a function that fairly estimates the LOS a given patient, admitted in a Psychiatric Department for acute patients, should have had according to his or her characteristics.

Age and gender relationship to psychiatric LOS have been reported in several studies [9, 12, 14, 16, 21]. Oiesvold et al [14] report longer LOS for patients in the older ages and for females in psychiatric patients in hospitals of Sweden and Finland. Huntley et al [12] classified age as one of the five variables significantly predicting LOS steadily over time. Barnow et al [9] found a correlation coefficient as high as 0.73 for describing the univariate association between age and LOS for depressed patients. Richter [21] found diagnosis and age were responsible of 10.5% of the LOS variations. Our results agree with these authors with regard to age but not to gender. Age is a natural determinant of LOS since it is closely related with all vital events; some authors (vg. Kiesler et al [7]), mix it up with another demographic variables while others like Tucker and Brems [8] just include it as a covariate.

Diagnosis is also a variable related to LOS in Psychiatric patients but how to include it with the aim of predicting LOS, is a challenge. Diagnosis Related Groups (DRG's) have been deeply explored [7, 22, 23]; they have the advantage of being just a few for Psychiatry though some authors have alleged they are not relevant for predicting LOS [24, 25]. The Diagnostic and Statistical Manual of Mental Disorders (DSM) classifications (III and IV, lately) [8, 13] or International Classification of Diseases (ICD 9^th or 10^th) [11] are also used in this context. Perhaps a broad classification system would achieve the best predictions but it would imply a huge number of patients for deriving the prediction function. We used an ad hoc classification based on DSM IV that yielded differences in LOS when analyzed univariately and when it was adjusted for other variables as well.

Most of the authors find an association between LOS and diagnosis [11‐14, 26] but there are some discrepancies, most authors report psychoses as responsible for the highest LOS [12, 14, 16] but others find major depression [13] as more important predictor of LOS.

It is recognized that the patient's severity of illness influence LOS independently from diagnosis [9, 17, 27] but finding a valid, reliable and useful way to measure it, with the aim of adjusting quality indicators, has always been and continues to be a challenge [28]). Various scales for measuring severity of illness in Psychiatric inpatients have arisen in the last two decades. The Brief Psychiatric Rating Scale (BPRS) in its expanded version [29], the Psychiatric Severity of Illness Index (PSII) [27], the Computerized Psychiatric Severity Index (CPSI) [30] and the Health of the Nation Outcome Scale (HoNOS) [31, 32] are probably the most mentioned ones. However, a low reliability is to expect during their use since appraisers must categorize symptoms in various levels of severity according to their opinion. Thus, implementation of any of these scales implies a period of special training and/or detailed instructions, a fact that prevents their use in daily practice. For instance, Durbin et al [33], in an attempt to introduce CPSI for predicting LOS from clinical records, had evaluators participate in a 3 day training program.

Our principal component solution is a real possibility since we collected information from all 19 symptoms described in the psychiatric record routinely used in our wards and afterward converted them in 8 factors. However, we still had to implement 3 categories for 8 of the symptoms (see Appendix), a feature that should be changed in the near future for the sake of gaining reliability.

We have not found any other study that employs PCA to reduce dimension with the aim of predicting LOS but the method has already been used to reduce symptoms' dimension in the field of Psychiatry and Psychology [34‐36].

ECT was also an important LOS predictor in our series. It has been included by other authors in LOS prediction models [11] or mentioned as a cause for longer stays [13, 37]. The use of ECT during hospitalization is a severity indicator, but is also a cause of complications.

Response to treatment turned out to be a variable with strong influence on LOS. Among reviewed literature only Draper and Luscombe [13] recognize the role of this issue in LOS prolongation. We understand it is a difficult aspect to assess and include in information systems unless the physician in charge of the patient completes the discharge form, a claim that should be evaluated in future research.

Other variables included in our study have been also explored by other authors. Marital status (or as living alone) has been acknowledged as an important LOS determinant in different studies [10, 11, 13, 16, 17, 38]; physical co morbidities have also been analyzed [13, 33, 39] and found fairly relevant. Place of residence, as distance from home to hospital was also included in a Brazilian study [40]; medications were also examined by Parks [10] and found "polypharmacy" as a LOS predictor in geropsychiatric patients.

About the goodness of fit of the regression function, 0.37 is not an encouraging determination coefficient but the difficulty to find functions that explain more than 50% of LOS variation, a variable of complex nature, is also true. Among the psychiatric domain, Creed et al were able to explain up to 49% of LOS variation including demographic data, clinical features, social measures and behavioural issues [11]. Richter found an R2 of only 0.11 including in its function: age, diagnosis and other clinical and sociodemographic variables [21]. Stoskopf and Horn found coefficients in the range of 0.10–0.14 including only diagnosis and severity [30]. Huntley et al achieved to explain 17% of total variance in LOS including five variables in a stepwise regression analysis [12]. Regarding the fitted model, we chose the logarithmic transformation of LOS since its original distribution was right tailed. LOS distribution has been explored by various authors. Priest et al analyzed LOS distribution in an acute Psychiatric department in London; he found the exponential model yielded the best fit [41]. Stevens et al fit the exponential model and explore the influence of several factors by means of a Cox Regression model, an approach that would not allow LOS prediction [16]. However, several authors [11, 13, 18] choose the logarithmic transformation for the search of predictors via a regression function, and perhaps most authors fit the regression model with the original LOS observations [10, 12, 17, 23, 33]. Marazzi in a huge European study found Lognormal, Weibull, or Gamma models were fine for describing the distribution of length of stay [42].

Regarding the proposed indicator (OLOS-PLOS), we confirmed the tendency of observed LOS to be higher than predicted LOS when there are inefficiency problems. However, we did not achieve a highly sensitive and specific cut-off point for detecting inefficiencies, a fact that emphasizes the necessity of refining the method with more variables and larger samples. A control method for efficient care similar to the proposed here is reported in some studies but not for psychiatric areas [43, 44].

It is fair to recognize that the process of detecting inefficiencies in the records was somewhat arbitrary; first of all we almost identify inefficiencies as delays though it could be argued inefficient care can be provided without any delay. However, it would be out of the scope of the study to search for another kind of inefficiencies, as, for instance the ones arising from a wrong management of the patient. In second place, some cut offs for deeming a record "inefficient" are also arbitrarily chosen; we chose the time intervals considered normal for the hospital usual performance in all departments. Perhaps the main limitation of the present approach is that the prediction function must be estimated in the same setting where allegedly inefficiencies exist. This handicap is partly solved with the elimination of outliers during the function development. Finally, a practical limitation of our method ensues from the many variables that must be reported by assistant physicians or retrieved from the records. We believe this issue can be solved with the introduction of computers and friendly computer programs at the wards.

Our work supports the importance of a series of variables as LOS predictors in a Psychiatry department. The observed-predicted length of stay can be implemented as an indicator of inefficiencies provided the appropriate cut-off point is chosen. The approach showed its validity and is adaptable to other settings although there is an obvious need to continue the effort in the search of more explanatory functions.