Early evaluation of patient risk for substantial weight gain during olanzapine treatment for schizophrenia, schizophreniform, or schizoaffective disorder

Post-hoc analyses were conducted using data from a large, randomized, controlled clinical trial within the Eli Lilly and Company database, and repeated using a smaller trial from the same database to test whether results were reproducible. In both studies, patients were diagnosed with schizophrenia, schizophreniform, or schizoaffective disorders; and weight was assessed weekly for the first 4 weeks of treatment and again at 28–30 weeks. For both studies, all patients provided informed written or witnessed oral consent after the study was described to them.

Study 1 was a 52-week trial comparing olanzapine (5–20 mg/day) and haloperidol (5–20 mg/day), conducted from June 1993 to February 1995 [17]. Of the 1336 patients who were randomized to treatment with olanzapine, 669 remained in the study and had weight recorded at approximately Week 30 (occurring between 26 and 34 weeks after initiation of study drug due to variation in visit intervals allowed by the protocol); of these, 661 had height data available allowing for calculation of BMI. The analysis population study included patients who were initially randomized to and continuously treated with olanzapine, up to 30 +/- 4 weeks, either within the double-blind arm or after being switched to open-label olanzapine rescue.

Study 2 was a 28-week, randomized, double-blind trial comparing olanzapine (10–20 mg/day) and risperidone (4–12 mg/day), conducted from April 1995 to January 1997 [18]. Of the 172 olanzapine-treated patients, 102 had a recorded weight at endpoint; of these, 101 had height data available, allowing for calculation of BMI.

These analyses examined 4 different definitions of substantial weight gain at endpoint (5, 7, 10 kg, and 7% of baseline weight) as well as 3 different definitions of substantial BMI increase (1, 2 and 3 kg/m²). Two different endpoint times (Week 30 for Study 1 and Week 28 for Study 2) were used due to the differences in scheduled visit intervals. These time points were chosen in order have comparable endpoints for both studies and because they are within the 21- to 39-week time period when weight gain associated with olanzapine treatment has been reported to plateau [19, 20].

Early weight gain at Weeks 1, 2, 3, and 4 was evaluated for usefulness in predicting risk of substantial weight gain. Early increase in BMI at Weeks 1, 2, 3, and 4 were evaluated for usefulness in predicting risk of substantial increase in BMI. Baseline variables examined as potential predictors of weight gain were age, Caucasian race, sex, smoking status and body mass index (BMI; kg/m²) [21].

The majority of analyses were performed for both weight gain (using early weight gain to predict substantial weight gain at endpoint) and BMI increase (using early BMI increase to predict substantial increase in BMI at endpoint). Analyses are explained below using weight gain as an example, but analogous analyses were also performed examining increase in BMI. Prediction rules for substantial weight gain were constructed using linear discriminant analysis (LDA, implemented using SAS^® PROC DISCRIM) for several sets of predictors: (i) baseline characteristics alone, (ii) baseline characteristics and weight change from baseline at specific early time points (Weeks 1, 2, 3, and 4), (iii) early weight change as a single predictor. In LDA, a prediction rule is based on classification score (or linear discriminant function) calculated for discriminating between groups (for instance patients with and patients without substantial weight gain). The classification score is based on the sum of scores for each individual predictor evaluated for every patient by multiplying each predictor's value by its associated coefficient in the linear discriminant function. This total score is compared with a threshold value to differentiate between patients predicted to and predicted not to experience substantial weight gain. The coefficients for individual predictors are estimated from the data so as to achieve the best discrimination between the two groups (in terms of the ratio of between-to-within group variation) using all predictors in the model. The threshold is estimated to ensure that patients are allocated to the group for which they have the largest probability of membership. When using early weight change as a single predictor, the threshold simplifies to the midpoint between the average early weight change in groups with and without substantial weight gain.

In using predictive models with multiple predictors, the classification score is based on a combination of the amount of early weight change and other variables included in the model. As a result, there is no single cut-off point for early weight gain that can be used to predict whether or not a patient will experience substantial weight gain. The threshold applies to the entire linear combination of multiple predictors, which implies that for the same amount of early weight gain, patients may be predicted to gain substantial weight or not gain substantial weight depending on the values for other characteristics included in the prediction model.

The performance of a classification rule can be evaluated by examining its operating characteristics (sensitivity, specificity, and positive and negative predictive values). To calculate these, the analysis population is divided into four groups:

∘ Group A–patients for whom the rule correctly predicted substantial weight gain;

∘ Group B–patients for whom the rule incorrectly predicted substantial weight gain (false positive);

∘ Group C–patients for whom the rule incorrectly predicted no substantial weight gain (false negatives); and

∘ Group D–patients for whom the rule correctly predicted no substantial weight gain.

The operating characteristics of a classification rule can then be calculated:

To evaluate the direction in which patient drop out may have affected characteristics of our analysis population (subjects available at week 30) we constructed logistic regression models for probability that a patient was included in the analysis population using as predictors patient's characteristics available at baseline and changes in psychiatric symptoms and weight at Week 3.

To examine how patient discontinuation could have affected the early weight gain cut-offs reported here, we performed a sensitivity analysis using multiple imputation [22, 23], imputing missing values for 30-week weight data for Study 1 patients who were available up to at least Week 3. Each missing value was imputed 100 times resulting in 100 completed datasets. The imputation was based on Bayesian regression that imputes missing observations for every patient using all available data (i.e., all repeated measures on weight up to the patient's discontinuation from the study). The imputation takes advantage of correlation in repeated measures; however, it also explicitly incorporates uncertainty in missing values about the expected trends (i.e. it does not simply impute future weight as values predicted from a linear regression on the previous values). As a result, multiple completed datasets were generated, mimicking the variability expected in missing observations. This procedure automatically adjusted for bias in observed data induced by the selection mechanism to the extent that the variables driving the selection bias were utilized in the imputation procedure. Imputation obviously cannot make up for any bias explained by variables that were not available or not included in the imputation model. Using imputation, we completed 100 datasets; each contained data on N = 1156 patients (number of olanzapine-treated patients available at Week 3). We evaluated the cut-offs for early weight gain and associated operating characteristics for each of these 100 datasets; summaries over 100 datasets (means, max and min values of operating characteristics) were reported.

Detailed demographics and other information on Studies 1 and 2 have already been published [16‐18, 21]. Briefly, olanzapine-treated patients in both studies were predominantly male, predominantly Caucasian with similar mean ages, and mean baseline BMIs (Table 1). In Study 1, lower BMI, lower age, male sex, non-Caucasian race and weight gain at Weeks 1 through 4 were all significantly correlated with later weight increase. In addition, lower BMI, younger age and weight gain at Weeks 1 through 4 were significantly correlated with gaining at least 7% of baseline weight. Endpoint for Study 1 in these analyses was 30 weeks, with a protocol allowed variation of 4 weeks. In Study 2, which had a smaller patient population, lower BMI and weight gain at Weeks 2 through 4 were significantly correlated both with weight change at endpoint and with gaining at least 7% of baseline weight by Week 28 (Table 1). Discontinuation due to weight gain as an adverse event was rare in both studies, occurring in less than 1% of olanzapine-treated patients.

In Study 1, change in weight at 30 weeks among olanzapine-treated patients ranged from a decrease of 22 kg to a gain of 27 kg; mean change was a gain of 5.4 kg (Figure 1A). In Study 2, change in weight at 28 weeks ranged from a decrease of 13 kg to a gain of 20 kg; mean change was a gain of 5.2 kg at Week 28 (Figure 1B). Magnitude and variability of weight change was similar between studies.

In Study 1, change in BMI at 30 weeks among olanzapine treated patients ranged from a decrease of 8.5 kg/m² to an increase of 10.8 kg/m²; mean change was an increase of 1.9 kg/m² (Figure 1C). In Study 2, change in BMI at 28 weeks ranged from a decrease of 5.3 kg/m² to an increase of 6.8 kg/m², mean change was a gain of 1.8 kg/m² (Figure 1D).

Tables 2 and 3 present operating characteristics for different definitions of substantial weight gain (Column 1) and predictors included in the model (Column 2) based on prediction rules derived from Study 1. Two sets of estimated operating characteristics are reported: the first was derived from Study 1 data (via leave-one-out cross-validation) and the second was obtained by applying the rules derived from Study 1 data to the validation data from Study 2. Also shown are the number and percent of patients who met the conditions of the prediction rules and who experienced the defined level of substantial weight gain or substantial BMI increase at the endpoint. In general, sensitivity and specificity improve as time of early weight/BMI assessment increases, indicating that predictions made based on data from Weeks 3 or 4 have greater discriminating power than those based on data from earlier time points.

Figure 2 shows specificities and sensitivities for Study 1 when substantial weight gain is defined as at least 10 kg at Week 30, allowing comparison of the discriminating ability of predictive models depending on time point for early weight gain and inclusion of baseline characteristics in the predictive model. For all measures, the accuracy improves as time of prediction moves from Week 1 to Week 3 and then appears to level off. In addition, inclusion of baseline characteristics have little or no effect on specificity, suggesting early weight gain without baseline characteristics may be sufficient for predicting which patients will not gain substantial weight. For sensitivity, once early weight gain is included in the prediction model, the additional predictive value of baseline characteristics diminishes and becomes marginal after 3 weeks.

While sensitivity and specificity are important for assessing discriminatory power of prediction rules, positive and negative predictive values (PPV and NPV) may be of more practical interest to clinicians interested in predicting how likely a given patient is to gain substantial amount of weight. Both PPV and NPV tend to increase at later time points, but the greater the defined level of substantial weight gain or BMI increase, the greater the NPV and the smaller the PPV. This is not surprising, as it is easier to predict a more common event (gain of ≥ 5 kg) than a less common one (gain of ≥ 10 kg).

As would be expected from the prediction rule when substantial weight gain is defined as a gain or 10 kg or more, based on a cut-off of weight gain at 3 weeks (Table 2), weight change distribution at endpoint for patients in Study 1 who gained less than 2 kg by Week 3 differs substantially from that of patients who gained 2 kg or more (Figure 3). Weight loss during treatment was much more common among the first group, and weight gain of as much as 25 kg appears to be more common among the second group.

These results suggest that the weight gain observed in our analysis population at Week 30 may be somewhat larger than the gain that would have been observed if all patients available at 3 weeks had remained in the study up to Week 30. However, from this observation alone it is not clear whether selection bias in weight change for our analysis population could have impacted our estimated cut-offs for early weight gain as predictors of substantial weight gain, and associated operating characteristics (sensitivity, specificity, PPV, NPV).

To evaluate whether patient discontinuations could have biased the results, we performed a sensitivity analysis using multiple imputation [22, 23], imputing 30-week weight data for patients who were available up to at least Week 3 and generating 100 completed datasets. Reported below are summary characteristics computed across these 100 datasets that reflect the hypothetical results that would be expected had no dropouts occurred; the difference between maximum and minimum values for the completed datasets reflect uncertainty associated with missing data. Mean change in weight from baseline to endpoint was 4.5 kg (minimum: 4.1 kg, maximum: 5.1 kg), somewhat lower than in the observed data. Correlations between weight change at Week 3 and weight change at endpoint across completed datasets, mean r = 0.50 (minimum: 0.43, maximum: 0.56), were somewhat higher than correlations in the observed data (r = 0.46) suggesting that dropouts could have caused some dilution of the relationship between early and late weight change in the observed data. Cut-off values for early weight gain used in predicting substantial weight gain at 30 weeks were similar to those obtained with the observed data. For example, for predicting weight gain of 10 kg, mean cut-off for early weight gain at Week 3 was 1.9 kg (minimum: 1.8 kg, maximum: 2.1 kg). For weight gain of 7% the mean cut-off for the percent weight change at Week 3 was 1.9% (minimum: 1.8%, maximum: 2.0%). Operating characteristics for weight gain of 10 kg with early weight gain at Week 3 were: mean sensitivity = 66.1% (minimum: 61.8%, maximum: 70.1%), mean specificity = 70.1% (minimum: 67.5%, maximum: 75.4%), mean PPV = 37.5% (minimum: 32.8%, maximum: 43.1%), and mean NPV = 88.4% (minimum: 85.7%, maximum: 90.6%). These values are all fairly close to those obtained using the analysis population. Thus, the selection bias in our analysis population had little effect on our results.