Secondary prevention of myocardial infarction in people with dementia: a multi-jurisdictional retrospective cohort study in Asia, Europe and the USA.

Cognitive behavioral therapy (CBT) is a first-line treatment for mild-to-moderate depression and anxiety disorders []. To expand treatment coverage, enhance flexibility, and reduce costs, CBT is increasingly delivered remotely. Existing evidence suggests that internet-delivered CBT (ICBT) is comparable to face-to-face CBT in reducing symptom severity, expands access to mental health services by mitigating patients’ financial and time constraints, and requires less therapist time, thereby increasing patient throughput and lowering health care costs [-]. However, similar to traditional CBT, up to 50% of patients do not experience clinically significant improvement [-] and are at risk of clinical deterioration and poor long-term outcomes []. Numerous studies have sought to identify robust predictors of treatment outcome to facilitate patient stratification. Existing evidence suggests several clinical (baseline symptom severity [,-] and comorbidity [,,]) and sociodemographic (education [,] and employment [,,]) predictors of CBT outcome; however, no reliable molecular or neuroimaging biomarkers have yet been identified []. Recent advancements in psychiatric genomics support a substantial contribution of genetic differences to the variance in complex traits, including the onset and prognosis of psychiatric disorders []. Genome-wide association studies (GWAS) provide data on population-level associations of single nucleotide polymorphisms (SNPs) and phenotypes of interest. SNPs are commonly aggregated into polygenic scores (PGSs), which summarize small effects of multiple risk alleles on a specific trait. Therapygenetics is an area of research that specifically aims to quantify the contribution of genetic variation to psychotherapeutic treatment outcomes []. Yet, the 3 existing GWAS of CBT response did not identify any genome-wide significant loci and failed to derive stable SNP-based heritability estimates, possibly due to insufficient sample sizes (n=980‐3113) [-]. However, there are some indications that common genetic variants may be implicated in CBT response variability. Preliminary findings include a positive association between PGS for educational attainment and symptom reduction [], a negative association between PGS for autism spectrum disorder and symptom reduction [], and a weak predictive effect of PGS for depression and intelligence on remission []. Despite some advances in identifying group-level predictors, it remains unclear whether they can be effectively translated into meaningful predictions at the individual patient level. Machine learning (ML) has increasingly been applied to predict a future outcome for a yet unobserved individual patient by first training a model using the abundance of retrospective data from other patients. Given evidence that models trained on multiple data types typically outperform single-modality models [], it has been recommended that future efforts focus on multimodal prediction [,]. The premise that justifies the deployment of predictive models in routine psychiatric care is that they add value beyond clinical judgment, which has been shown to be affected by bias and overoptimism [,]. Due to a growing interest in precision medicine, new prediction tools emerge continuously; however, most are not adopted clinically, mainly due to low accuracy and lack of external validation. Systematic reviews and meta-analyses of ML studies predicting treatment outcome [-] report varying results but emphasize that insufficient sample size lies at the core of inflated performance and poor generalization. In their comprehensive review, Sajjadian et al [] highlight a strong negative relationship between study quality, defined by sufficient sample size and robust validation methods, and predictive accuracy. The authors raise a concern that the overly optimistic results reported by most reviewed papers stem from insufficient methodological scrutiny rather than genuinely high predictability. Finally, implementation studies applying and assessing predictive models in real-world clinical practice are scarce, demonstrating a substantial translational gap [,]. In summary, despite the promise of ML approaches, translation to reliable individual-level predictions has been hampered by several persistent gaps: small sample sizes, single data types, and methodological flaws.