In this work, we implemented group-specific models of whole-brain CTh according to chronological age for HC and YDD for AD and FTD patients using SVR. We modeled CTh changes with time, and we demonstrated these models’ capability to identify individual variations as deviations from the norm. We studied individual deviations from the model defined for HC and from the disease-specific models. In AD, individual deviations from the HC-defined model correlated with YDD and MMSE scores. Using the disease-specific models, we found a significant inverse correlation between CSF-NfL levels and the deviation from the model in most of the brain regions for FTD patients, providing additional evidence of the relationship between imaging and fluid biomarkers in FTD. Furthermore, we found a positive correlation for AD between the MMSE scores and the deviation from the model in multiple regions. Finally, using real longitudinal data, we performed an exploratory study to evaluate the ability of the models described above to obtain individual predictions of future CTh values.

In healthy subjects, our model estimated whole-brain CTh using the chronological age of the subjects with residual values near zero, indicating a good fit. Several studies and multi-centric initiatives have recently focused on studying normative models with age in healthy samples [15]. Even if our healthy sample is limited compared to these studies, our results align with what has been described. In addition, we offer the applicability of such models with longitudinal data with satisfactory results. Furthermore, as deviations from normality have been widely used to assess disease severity and cognitive variability in youth [28] and in psychiatric disorders [29], we study if AD and FTD patients could be described with the previous HC model. As expected, individual AD and FTD CTh data diverged from the model defined with HC subjects. In AD, this divergence was significantly correlated with YDD and MMSE but not in FTD, suggesting that both time and general cognition impact more consistently in CTh changes compared to HC in AD. In AD, the regions of the temporal, parietal, and occipital lobes were the ones with the highest deviations from the HC model. These regions have been reported to be the most affected in patients with AD [30,31,32,33]. Therefore, our study can identify meaningful atrophy patterns using estimation models. On the other hand, the temporal and frontal lobes were the ones with the highest deviations for the FTD patients compared to the HC trend. In addition, temporal regions presented higher deviation than the AD group, depicting higher structural alterations in FTD [11, 32, 34, 35].

Given that models derived from the HC group that used chronological age showed high variability for these groups, we expected that modeling disease duration could better capture these variations. In this sense, we described CTh models in AD and FTD using YDD instead of the chronological age even though both AD and FTD individual values evidence the existence of variability within these diseases. These results complement other research highlighting these models to assess heterogeneity in neurodegenerative disorders [1, 36, 37].

We found that the individual AD deviations from the AD CTh model in the right precentral and right entorhinal inversely correlated with the NfL levels. At the regional level, the correlations between MMSE scores and the deviations from the model in AD were in the temporal and parietal lobes. In AD, MMSE has been associated with CTh [4, 33, 38]. In our case, the AD patients who deviated more from the model trend had high MMSE scores. Previous studies such as Valtteri et al. [38] showed that less CTh in frontal and temporal lobes is associated with lower MMSE scores in the AD group, which coincides with our regions and the direction of the correlation. Even if we study variability and not direct CTh measures, the regions that appear significant in the correlation analysis have high coincidence with the AD cortical pattern, which includes the temporal and parietal lobes appear [30, 31, 33, 34].

The deviations of the FTD patients from their own model presented correlations with CSF-NfL levels in the frontal, temporal, and parietal lobes. The frontal and the temporal areas are known to be the most affected in patients with FTD compared to HC [32, 34, 39]. In FTD, NfL levels have been suggested as a maker of the disease severity [40, 41]. We found that higher levels of NfL corresponded to lower residual values. Thus, patients with CTh patterns that adjusted well to the model had higher CSF-NfL levels, suggesting that the model captures the trend of neurodegeneration. Instead, lower NfL levels could correspond to patients that move further away from the FTD model, as has been described for slowly progressive FTD or non-progressors FTD [42]. Indeed, when studying the different FTD variants, we found that the main pattern was mainly driven by bvFTD, as other variants, such as PPA, showed much higher deviation. In addition, the fact that the parietal lobule emerged, which has not been reported previously, could be due to the high presence of bvFTD patients in the FTD sample. At the same time, the individual deviations in PPA patients did not present any significant correlation with NfL, suggesting that the heterogeneity in these patients' CTh pattern is not related to CSF-NfL levels. These results might provide additional support to the use of NfL levels as a marker of neurodegeneration and disease severity in FTD, paving the way for its future use as an outcome measure for clinical practice. In addition, the inexistent correlation with MMSE scores in FTD patients could be due to the fact that MMSE is not a sensitive cognitive measure in these patients, contrary to AD patients.

At the longitudinal level, we explored the application of these predictive models, trained at a cross-sectional level, to predict CTh patterns for future visits. We used the longitudinal data to know if the pseudo-longitudinal models developed with baseline data could predict the CTh values of future visits of these patients. We compared the predicted values of the longitudinal data with the real values of those visits. For healthy subjects and AD, we found that the model could predict well the CTh values at future visits. For FTD patients, the variability observed at baseline was reproduced for the longitudinal data and increased at each visit. Then, we focus on the FTD variants; the bvFTD patients fit best in future prediction model, and the nfvPPA patients were the worst. Our results align with previous studies showing that FTD is phenotypically heterogeneous [21, 22] and support that an independent predictive model should be created for each FTD phenotypic variant. Similarly, in a recent study of genetic FTD, Poos et al. [43], identified differences between the three major FTD gene mutation carriers' trajectories, suggesting that each genetic FTD subtype should also be modeled separately. The fit showed more FTD heterogeneity than AD at baseline and longitudinal levels. Overall, the proposed models could be the starting point to be able to differentiate different subtypes or dementias through the estimation of their CTh values or variabilities. Thus, further studies with a bigger cohort should study this point in detail.

Our study has some limitations. One of the major limitations of the study is the sample size. The sample size limited the generation of different CTh models for each FTD variant, especially all models at the third visit. However, the unicentric nature of the study, even if limited in the sample size, also provided homogeneity to the data acquisition. We only tested the effect of two selected neurodegeneration biomarkers in the individual residuals from their disease model. The selection of these two biomarkers was, at a certain point, arbitrary. Future studies could evaluate whether other markers could better explain the individual variability.

In conclusion, SVR provides the opportunity to generate CTh disease models to predict longitudinal changes and to study individual variability in AD, FTD, and healthy individuals and their relationships with cognitive features and biomarkers.