Predicting health and socioeconomic outcomes in patients with common psychiatric disorders

<jats:p>&lt;p dir="ltr"&gt;Major depressive disorder, panic disorder, and social anxiety disorder are among the most common psychiatric conditions, closely tied to socioeconomic disadvantage and to substantial impairment in functioning and work. Cognitive behavioral therapy (CBT), including its scalable internet-delivered form (ICBT), is a recommended first-line treatment, yet up to half of patients do not achieve a clinically meaningful symptom reduction and a minority even deteriorate. Tools to reliably predict which patients are at risk of poor health and socioeconomic outcomes are largely lacking, which limits the ability to tailor care and allocate scarce resources efficiently. The aim of this thesis was to investigate whether health and socioeconomic outcomes after ICBT can be predicted from multimodal data - clinical, sociodemographic, register-based, treatment-course, and genetic - using traditional statistical, quantitative genetic, and machine learning approaches.&lt;/p&gt;&lt;p dir="ltr"&gt;&lt;b&gt;Study I&lt;/b&gt; drew on the MULTI-PSYCH cohort of 2,668 patients treated with guided ICBT for depression (n = 1,300), panic disorder (n = 727), or social anxiety disorder (n = 641) at the Internet Psychiatry unit in Stockholm between 2008 and 2020, linked to nationwide Swedish registers and genotyped. Two linear regression models were compared: a baseline model with six predictors collected during pre- treatment screening and a full model adding register-based predictors and seven polygenic scores for psychiatric disorders and cognitive traits. The baseline model explained 27% of the variance in post-treatment symptom severity and the full model 34%. Alongside baseline severity and socioeconomic disadvantage, the analysis identified previously understudied associations, notably with comorbid autism spectrum disorder and attention-deficit/hyperactivity disorder, whereas polygenic scores provided no meaningful additional explanatory value.&lt;/p&gt;&lt;p dir="ltr"&gt;&lt;b&gt;Study II&lt;/b&gt; performed a genome-wide association study (GWAS) of symptom change following CBT and estimated single nucleotide polymorphism (SNP) heritability. To assemble the largest such sample to date (N = 3,113), the MULTI- PSYCH cohort was combined with the patients treated for obsessive-compulsive disorder (OCD) from the Nordic OCD &amp; Related Disorders Consortium (NORDIC). No genome-wide significant variants were identified, and SNP-based heritability of symptom change was estimated at h3Np = 0.221 (SE = 0.123), consistent with a possible modest contribution of common genetic variation. Current samples nevertheless remain underpowered for genetic discovery.&lt;/p&gt;&lt;p dir="ltr"&gt;&lt;b&gt;Study III&lt;/b&gt; used the MULTI-PSYCH data to develop and temporally validate machine learning models of varying complexity for prediction of clinically meaningful improvement after ICBT. Moderate performance was achieved across algorithms (AUCtest 0.732-0.749). Models incorporating register data generally outperformed the benchmark model based on pre-treatment screening data, whereas polygenic scores added no incremental value.&lt;/p&gt;&lt;p dir="ltr"&gt;&lt;b&gt;Study IV&lt;/b&gt; leveraged the same cohort and machine learning framework to predict labor market marginalization (disability pension, long-term sickness absence, or long-term unemployment) one year after ICBT and evaluated whether treatment- course data improved performance. The results showed moderate discrimination (AUCtest 0.754-0.768) but limited sensitivity (37-40%). Prior work disability, functional impairment, and long-term psychiatric morbidity dominated the predictive signal. Treatment-course data contributed little beyond baseline predictors, and polygenic scores added no incremental value.&lt;/p&gt;&lt;p dir="ltr"&gt;Taken together, clinical and socioeconomic outcomes after ICBT were predictable to a moderate degree, with the strongest and most consistent signal arising from baseline clinical, functional, and socioeconomic burden. Richer multimodal data explained more variance in post-treatment symptom level than the sparse set of routinely collected predictors and uncovered several novel markers of poorer outcome. Register linkage contributed a modest improvement in predictive performance. Common genetic variants, by contrast, provided consistently limited explanatory and predictive value: the GWAS of symptom change detected no significant loci and yielded only an imprecisely estimated SNP-based heritability, and polygenic scores added no incremental value in any of the models.&lt;/p&gt;&lt;p dir="ltr"&gt;Prediction of clinically meaningful improvement reached moderate discrimination and warrants prospective clinical validation. Prediction of future labor market marginalization showed limited sensitivity and positive predictive value but may still support cautious risk stratification of patients with a greater need for coordinated vocational and social support. Short-term symptom improvement and later labor market participation appeared to reflect partly distinct prediction targets. The constituent studies thus show that multimodal prediction in routine psychiatric care is feasible and, subject to prospective validation, could help identify patients in need of closer monitoring and additional support, providing a foundation for responsible development of prognostic tools in precision psychiatry.&lt;/p&gt;&lt;h3 dir="ltr"&gt;List of scientific papers&lt;/h3&gt;&lt;p dir="ltr"&gt;I. &lt;b&gt;Kravchenko O,&lt;/b&gt; Bäckman J, Mataix-Cols D, Crowley JJ, Halvorsen M, Sullivan PF, Wallert J, Rück C. Clinical, genetic, and sociodemographic predictors of symptom severity after internet- delivered cognitive behavioural therapy for depression and anxiety. BMC Psychiatry. 2025 May 30;25:555. &lt;a href="https://doi.org/10.1186/s12888-025-07012-x" target="_blank" rel="noreferrer"&gt;https://doi.org/10.1186/s12888-025-07012-x&lt;/a&gt;&lt;/p&gt;&lt;p dir="ltr"&gt;II. Bäckman J, &lt;b&gt;Kravchenko O,&lt;/b&gt; Halvorsen M, de Schipper E, Ivanova E, Kaldo V, Hentati Isacsson N, Olsen Eide T, Höffler KD, Mattheisen M, Hansen B, Kvale G, Hagen K, Haavik J, Mataix-Cols D, Crowley JJ, Wallert J, Rück C, Nordic OCD and Related Disorders Consortium (NORDIC). Genome-wide association study of symptom change following cognitive behavioral therapy for common mental disorders. Am J Med Genet B Neuropsychiatr Genet. 2026;201(6):396-405. &lt;a href="https://doi.org/10.1002/ajmg.b.70015"&gt;https://doi.org/10.1002/ajmg.b.70015&lt;/a&gt;&lt;/p&gt;&lt;p dir="ltr"&gt;III. &lt;b&gt;Kravchenko O,&lt;/b&gt; Halvorsen M, Bäckman J, Kaldo V, Crowley JJ, Kuja- Halkola R, Rück C, Wallert J. Prediction of clinically meaningful improvement after internet-delivered cognitive behavioral therapy for depression and anxiety disorders: Machine learning-based predictive model development and temporal validation study. J Med Internet Res. 2026;28:e100162. &lt;a href="https://doi.org/10.2196/100162"&gt;https://doi.org/10.2196/100162&lt;/a&gt;&lt;/p&gt;&lt;p dir="ltr"&gt;IV. &lt;b&gt;Kravchenko O,&lt;/b&gt; Halvorsen M, Bäckman J, Kaldo V, Crowley JJ, Kuja- Halkola R, Rück C, Wallert J. Prediction of labor market marginalization in psychiatric patients following internet-delivered cognitive behavioral therapy. Research Square. 2026 Jun 9. [Manuscript Preprint] &lt;a href="https://doi.org/10.21203/rs.3.rs-9759052/v1"&gt;https://doi.org/10.21203/rs.3.rs-9759052/v1&lt;/a&gt;&lt;br&gt;&lt;/p&gt;</jats:p>