In this study, we present two key findings that shed light on the heterogeneity of PSA. Leveraging SCCA in conjunction with cluster analysis, we successfully deciphered PSA heterogeneity, revealing distinct biotypes characterized by unique clinical features. Additionally, through a combination of data-driven and hypothesis-driven approaches, we demonstrated that ERDN functional connectivity pattern variants play a central role in PSA heterogeneity, providing crucial insights into the neurobiological characteristics of these diverse biotypes through multi-scale fMRI parameters.
Similar to the results using multivariate analysis combined with clustering algorithms 9,17−19, we succeeded in classifying heterogeneous disorders into multiple biotypes. In our study, there were significant differences in scores on the dimensions of apathy across biotypes (Fig. 3e). This confirms the heterogeneity of the deconstructed PSA in the study and breaks with the traditional diagnostic framework, which tends to oversimplify syndromes with a single label.
We imparted clinical significance to the identified biotypes based on patients' clinical trajectories (Fig. 3d). Biotype 1, characterized by elevated depression and apathy scores at baseline, showcased no deviation from the overall apathy score at follow-up, earning its distinction as the 'depression leading to apathy symptom group.' In contrast, biotype 2 exhibited higher apathy scores at both baseline and follow-up, justifying its categorization as the 'apathy syndrome group.' Biotype 3, with lower apathy and depression scores at both baseline and follow-up, emerged as the 'relatively healthy group.' Lastly, biotype 4, marked by elevated depression scores at both time points, was identified as the 'depression syndrome group'.
Multiscale fMRI parameters of the ERDN reaffirmed and elucidated these distinctions, providing a multidimensional perspective of the neurobiological functioning within each biotype (Fig. 4c, d). Specifically, biotype 3 demonstrated above-average global functioning of the ERDN, suggesting relatively normal neurobiological functioning in this group. In contrast, biotype 2, the 'apathy syndrome group,' exhibited the lowest graph theoretic metrics scores among all the groups, indicating a significant link between ERDN network damage and the onset of apathy. Notably, while biotype 1's graph-theoretic indicator scores did not significantly differ from either biotype 3 or the overall mean, functional segregation metrics highlighted abnormal performance in crucial brain regions like the ACC and medial orbitofrontal lobe.
In clinical practice and research, apathy is often overlooked as part of depression, or it is difficult to distinguish apathy from depression using traditional methods.2. This poses significant challenges for both patient treatment and the interpretation of research findings. For example, some studies have found that antidepressant treatments may exacerbate symptoms of apathy20. And Lauren et al. found in elderly depressed patients that apathy may contribute to antidepressant side effects21. In our study, biotypes 1, 2, and 4 have similar symptoms but different or even completely opposite neurobiological bases. This discovery could explain why patients with similar symptoms in these studies had different treatment responses when receiving antidepressant therapy. Thus, our findings on PSA heterogeneity may have implications for future PSA research and clinical trial design.
Our other major finding is that the performance of RSFC for prediction using ERDN for some biotypes is similar to or better than RSFC for prediction using all ROIs. Although the predictive power of the ERDN model was reduced for biotypes 3 (relatively healthy group) and 4 (depression syndrome group), good performance for biotype 1 (depression leading to apathy symptoms group) and biotype 2 (apathy syndrome group) supports the idea that the ERDN is a central mechanism in the formation of apathy. In addition, The favorable predictive performance of the ERDN model for biotype 4 (depressive syndrome group) supports the recent view that there may be a common mechanism between apathy and depressive syndromes that cuts across traditional diagnoses 4. Our study underscores the importance of ERDN in PSA heterogeneity and its potential as a promising target for treatment. By focusing on modulating neural activity related to reward decision-making, more effective and specific treatments for PSA patients can be developed.
Acknowledging certain limitations in our study, the risk of overfitting in CCA analysis stands as a notable concern. To mitigate this, we implemented various strategies such as feature reduction, feature selection, and regularized CCA with finely tuned parameters. Although these measures effectively reduced overfitting for the first canonical correlation, caution is warranted when interpreting the second canonical correlations due to its increased susceptibility to overfitting. Furthermore, the linearity assumption of CCA may limit its applicability, as neural activity and clinical information may feature nonlinear relationships. In the future, exploring alternative approaches, such as nuclear CCA or deep CCA, may offer potential solutions, although larger sample sizes in collaborative multicenter studies may be necessary. Additionally, the current study utilized statistical inference to assess clustering, but it remains challenging to determine whether clustering is truly meaningful, even with rejection of the null hypothesis against multivariate Gaussian distribution. To address this, we suggest further investigations involving predicting treatment outcomes and long-term prognosis based on identified biotypes. Lastly, the lack of imaging follow-up data and reliance on unimodal imaging limit the neurobiological interpretation. Thus, future research should focus on multimodal multivariate analyses, utilizing large sample sizes from multicenter studies with longitudinal follow-up and treatment information to address these limitations effectively.
In conclusion, our study provides seminal insights into the understanding of PSA heterogeneity and its neurobiological basis. The critical role of ERDN in PSA heterogeneity was identified, providing promising avenues for future study design and individualized therapy development.