Natural history of retinitis pigmentosa based on genotype, vitamin A/E supplementation, and an electroretinogram biomarker

BACKGROUND A randomized clinical trial from 1984 to 1992 indicated that vitamin A supplementation had a beneficial effect on the progression of retinitis pigmentosa (RP), while vitamin E had an adverse effect. METHODS Sequencing of banked DNA samples from that trial provided the opportunity to determine whether certain genotypes responded preferentially to vitamin supplementation. RESULTS The genetic solution rate was 587 out of 765 (77%) of sequenced samples. Combining genetic solutions with electroretinogram outcomes showed that there were systematic differences in severity and progression seen among different genetic subtypes of RP, extending findings made for USH2A, RHO, RPGR, PRPF31, and EYS. Baseline electroretinogram 30-Hz flicker implicit time was an independent, strong predictor of progression rate. Using additional data and baseline implicit time as a predictor, the deleterious effect of vitamin E was still present. Surprisingly, the effect of vitamin A progression in the cohort as a whole was not detectable, with or without data from subsequent trials. Subgroup analyses are also discussed. CONCLUSION Overall, genetic subtype and implicit time have significant predictive power for a patient’s rate of progression, which is useful prognostically. While vitamin E supplementation should still be avoided, these data do not support a generalized neuroprotective effect of vitamin A for all types of RP. TRIAL REGISTRATION ClinicalTrials.gov NCT00000114, NCT00000116, and NCT00346333. FUNDING Foundation Fighting Blindness and the National Eye Institute: R01 EY012910, R01 EY031036, R01 EY026904, and P30 EY014104.

(The reference group, which would be the first or last group by convention, was chosen as "group 2" for historical reasons only.) These dummy variables allowed for the use of linear contrasts to model effects of vitamin A and E, as was done for the original study, and as is recommended for 2x2 factorial designs (1, 2). Sample SAS code is given below: proc mixed data=use; class year ; model log30HzERGOU = year_numeric group1 group3 group4 year_times_group1 year_times_group3 year_times_group4 /solution; random intercept /subject=person_id; estimate "A effect" year_times_group1 0.5 year_times_group3 0.5 year_times_group4 -0.5 ; estimate "E effect" year_times_group1 -0.5 year_times_group3 0.5 year_times_group4 0.5 ; estimate "interaction" year_times_group1 1 year_times_group3 -1 year_times_group4 1 ; The covariance matrix between study years was inspected and had smoothly decreasing values moving away from the diagonal with range 0.97-0.90, and therefore a random intercept model was used, rather than a repeated measures model with a structured covariance matrix.
For power calculations for longitudinal progression rates within gene subgroups, power was estimated using the sample size formula: 1− = √ * 2 2 * 2 − 1− 2 , where β is the power corresponding to the z-score of a normal distribution, N is the total sample size, is the effect size for the yearly rates of change, 2 is the variance of the estimated slope which equals the sum of between-subject variability and within-subject variability, α is the significance level which is calculated as 5%. The 2 in the denominator reflects the two comparison groups (with and without treatment), for similarly sized groups. Rates

Longitudinal analyses in R
Analyses were performed in version 4.1.3 using the Lmer package. Sample R code is below.
"Year" was the study year. "person_id" was the subject id. "Trt" was defined as a factor representing Gene-specific models were also created in R using a similar model. In the SAS analysis above and in the original study, the year variable had been rounded to the nearest integer. In the R lmer package, the rounding is not required. Rounded or unrounded year data gave nearly identical results (though small effects can cause crossing of the significance threshold for the borderline p-values, such as the EYS group -see Results). For the model using propensity score matching (Table 2), matching was implemented using the MatchIt package/function in R. For the outlier detection of baseline implicit time described in the text, boxplot.stats()$out was used to detect outliers.

Homogeneity testing
For homogeneity testing between the three clinical trial data sources, we assessed for homogeneity of the three data sets using the following linear regression model: Ln(ERG amplitude) = b0 + b1*treatment group + b2*year + b3*source + b4*baseline implicit time + b5*treatment group*year + b6*source*year + b7* baseline implicit time *year, where the b coefficients are fit by the regression model, and "source" is the variable representing one of three clinical trial data sources. The beta described in the text to determine homogeneity is b6. -9.6 ± 2.2 -9.2 (3) RHO -6.8 ± 1.5 -8.7 (4) RPGR * -9.0 ± 1.8 -7.1 (5) USH2A