The first goal of this study was to identify predictors of post-treatment symptom severity in patients with mild to moderate MDD, PD, and SAD treated with highly standardized ICBT protocols and using a large and diverse pool of clinical, genetic, and register predictors. Our findings support the previously established importance of pre-treatment symptom severity, psychiatric comorbidities, family history of psychopathology, and socioeconomic status. In line with existing research on the link between marital status and mental health, being single, divorced or widowed was predictive of a poorer treatment outcome. Furthermore, thanks to unique data sources available in this study, we were able to assess a variety of predictors that have not been investigated before: the role of prior and concurrent psychiatric diagnoses and specific medications is a novel addition of this study, as are socioeconomic predictors such as income and receipt of financial benefits.
Identification of comorbid ASD and ADHD as strong hindering factors bears potential clinical relevance. The need to adjust psychotherapeutic interventions for depression and anxiety to the patient’s ASD symptomatology has long been acknowledged. Socio-communication impairment, difficulties with introspection, and limited cognitive flexibility have been suggested as impeding treatment effectiveness72. Consequently, multiple CBT adaptations for patients with comorbid ASD have been developed73. The discovery that in this patient group, the average post-treatment score was nearly twice as high as for patients without ASD potentially highlights absence of relevant modifications in the studied ICBT treatment. In patients with comorbid ADHD, the findings may reflect a similar lack of treatment accommodation. Perhaps unsurprisingly, maintaining attention and organizing oneself over a 12-week treatment period, which involves extensive homework and the absence of immediate therapist guidance that is characteristic of ICBT, may pose a challenge for this patient group and calls for adaptive treatment strategies to avoid treatment failure74.
The application of PRS in psychiatry is promising as it is a constant throughout lifetime and can be used for long-term prediction. Yet, their limitation is that on a population level, PRS tend to follow a Gaussian distribution with significant overlap between cases and controls, and they only explain a small fraction of genetic variation, which in turn explains around 50% of phenotypic variation. Failure to detect the association between PRS and post-treatment symptom severity in our study is perhaps not surprising. We used PRS related to psychopathology and personality, and, while genetic liability to these traits is likely to be shared with that of treatment outcome, it does not necessarily follow that the same PRS will be helpful in predicting it. Thus, a specific PRS constructed from GWAS of treatment response, which is still missing, could potentially yield a stronger effect.
The second goal of the study was to assess whether employing a wider range of predictors would be advantageous in identifying patients who will have higher post-treatment symptom severity. We found that the full model with multimodal predictors had a superior performance in explaining the variance in post-treatment severity compared to the baseline model. However, the improvement in explained variance was modest. One potential interpretation is that even though the evaluated predictors encompass a distinct range of properties, they might, to a certain degree, be considered as capturing a shared underlying construct and providing little independent information. For example, genetic predictors will inevitably be correlated due to vertical pleiotropy, whereby the same genetic variant affects multiple different traits, e.g., a relevant SNP influences intelligence, which serves as a mediator influencing educational attainment, SES, and, in turn, post-treatment symptom severity. While the collinearity of independent variables does not negatively affect the explanatory power of the whole model, it does not improve it either, since many variables explain an overlapping proportion of variance in the outcome. Further, the full model would likely benefit from the inclusion of additional strong predictors. For example, in the exploratory complete-case analysis, duration of symptoms was strongly associated with the outcome in its effect size and variance explained but could not be included as it surpassed the predefined threshold for the allowed amount of missingness. Moreover, human complex traits are dominated by stochasticity and are often attributed to idiosyncratic factors which remain largely unmeasured. Another possibility is that pooling the three disorders makes the studied phenotype more heterogeneous and possibly dilutes the findings. Also, while the current sample size exceeds those in the earlier studies, it may still not be sufficiently powered. Finally, the relationship between predictors and the outcome may be non-linear and contain underlying interaction effects, making classical ordinary least squares modelling unsuitable.
Strengths and limitations
The main strength of this study is its relatively large sample size and diversity of evaluated predictors. Moreover, register-based data have almost no missing values and are highly reliable. Another strength is a homogenous group of patients that completed a highly protocolized treatment, which allows for meaningful comparisons of the outcome. Deriving the data from routine care has additional benefits since most studies of treatment outcome constitute a primary or secondary analysis of RCT data, where user characteristics are subject to stringent inclusion criteria, thus introducing selection bias and limiting the generalisability75. However, a potential limitation in the applicability of the findings to other populations needs to be mentioned. All the study participants lived in Sweden, were more educated than the general population, were fluent in Swedish, and exhibited mild to moderate symptom severity. In addition, there is a significant self-selection bias, with most patients self-referring to the Internet psychiatry clinic.
Finally, it must be emphasized that the findings of this study cannot be viewed as supporting the causal nature of the relationship between any of the predictors and the outcome. While the terms treatment outcome and treatment response do semantically assume, at least partially, that the post-treatment symptom measure is conditional upon and a direct consequence of the intervention, no such claims can be made given the non-experimental design of the study. However, causal language is widely and loosely applied across the literature and, in the interest of brevity and recognition, is also sometimes used in this paper, more so when referring to the previous findings rather than the current study, where we try to adapt the more observational term post-treatment symptom severity. This operationalization of the response variable was chosen as it is deemed by the authors as the most appropriate measure to be predicted at the baseline. Since the goal of the future predictive model is not to assess treatment effectiveness, relative symptom change was not of interest here. Moreover, even when the percentage change is substantial and passes some predefined threshold, the patient may still be symptomatic beyond what is intended. Choosing remission as the outcome was rejected due to its somewhat arbitrary cut-off, loss of information, and diminished statistical power that ensues from dichotomization.