gms | German Medical Science

Parkinson's disease (PD) is a common disease of the elderly [Ref. 1]. Due to its chronic character and age-dependent incidence, it poses a major burden for health security systems in most if not all western countries. Health economic evaluations are often seen as an aid for health policy decisions. These studies need a measure of health-related quality of life (QoL) to calculate utilities and quality-adjusted life years (QALYs). Utilities are not yet routinely collected in clinical trials. However, the Unified Parkinson's Disease Rating Scale (UPDRS) is a well-established and routinely used clinical scale which includes in several subscales items about activities of daily living (part II, "U2"), results of the physical examination of the patient (part III, "U3") and complications of the disease or therapy (part IV, "U4") [Ref. 2].

We have recently proposed an algorithm for the estimation of utilities based on EQ-5D values (EuroQol) from the sum scores of U2 - U4 of the UPDRS [Ref. 3]. In the present study we have set out to investigate, whether single items of the UPDRS may be sufficient to estimate utilities or may even improve our previous prediction rule.

We analyzed data from the 6-month's follow-up of prospective cost study (n=145) within the German Competence Network for Parkinson Syndroms (KNP) [Ref. 4]. In the dataset, data about EQ-5D index and UPDRS items 5-42 (corresponding to parts U2 - U4) were available. A classification and regression tree (CART) model was built to identify the most influential items on QoL as the outcome [Ref. 5]. As test statistic the t-test was used. Due to missing values, only 122 patients could be analyzed. Therefore, splits leading to 11 or less patients were not carried out (SQRT (122) = 11.05). Adjustment of the p-values of the individual tests of non-dichotomous variables followed the method of Miller and Siegmund [Ref. 6].

In a second step, the resulting splitting-variableswere used as predictors in linear regression models.

The CART-analysis resulted in a model with 3 levels. A first split was in variable U28, which evaluates the posture of patients on a scale from 0-4 leading to subgroups of 94 (group A) and 28 patients (group B). The group A was divided by item U27. This item contains results of the problems of patients with the task of rising from a chair. The division led to subgroups of n=25 (with no further possible splits) and n=69, which could be divided by item U36 (unpredictable fluctuations, i.e. sudden loss of drug effect leading to reduced mobility). The smaller subgroup of the first split, group B (n=28) was subdivided by item U22 (level of rigidity). The resulting 5 groups had mean (median) levels of QoL of 0.90 (0.89), 0.81 (0.89), 0.68 (0.70), 0.66 (0.70) and 0.32 (0.29), the overall QoL was 0.75 (0.89).

We analyzed several regression models, including only the items involved in the first two or all three levels, in raw or dichotomized form, or together with the respective interaction terms. Although not all predictors were significant on the 5%-level, the prediction as judged by adjusted R² was in the same magnitude as our previous model (0.49 to 0.55 as compared to 0.55 of our previous model). Models with the items U22, U27, with or without U28 and U36 showed the highest values of R², all being around 0.55.

We have presented preliminary results indicating that single items of the UPDRS may be as good for predicting QoL in PD patients as a prediction rule based on the sum of the subscores of the UPDRS parts II - IV. A better prediction than the previous prediction algorithm could not be shown. On the other hand, the items identified seem plausible from a clinical point of view, and reasonable subgroups have evolved. Tree-based prediction models can be advantageous for clinically oriented researchers. However, these results need validation and further research seems warranted.