In this study, we developed classifiers that predicted the conversion of patient status from Aβ (−) to Aβ (+) using baseline information on demographics and neuroimaging test results from the ADNI database. The major findings of this study are as follows. First, the incidence of conversion to amyloid positivity was 23.1% (53/229) within 5 years. Second, age, gender, and APOE genotype, but not family history, were effective in predicting the conversion to amyloid positivity. Finally, the prediction model, which consisted of age, gender, APOE genotype, and global SUVR, showed good accuracy (AUROC = 0.814). Furthermore, the addition of regional SUVR led to an improvement in the prediction performance (AUROC = 0.841). Our findings highlight the distinctive features that should be taken into consideration when selecting candidates for primary prevention treatment in CN individuals.
Our first major finding was that the incidence rate of conversion to amyloid positivity was 23.1% within 5 years, as 53 out of 229 participants were Aβ converters. Our findings are consistent with those of the previous studies. Specifically, recent studies from different cohorts have reported the annual incidence of conversion to Aβ (+) among elderly Aβ (−) CN individuals, ranging from 3.1 to 13% [29, 30]. Thus, approximately 20% of baseline Aβ (−) individuals converted to Aβ (+) status, suggesting that these individuals need to be considered candidates for primary prevention.
Age, gender, APOE genotype, and family history are well-known risk factors for amyloid positivity. However, in the present study, age, gender, and APOE genotype, but not family history, were predictive of conversion to amyloid positivity. Our findings are in line with previous studies reporting factors associated with the rate of Aβ accumulation, such as APOE genotype [29], age, and sex [16] in Aβ (−) individuals. However, our findings contradict a previous study, which reported a lack of obvious differentiating demographic features between amyloid converters and non-converters [30]. The discrepancy could be attributed to the differences in study designs between the reference and the present study, given the limited follow-up (median imaging follow-up was 1.3 years) and considerably smaller sample size of baseline Aβ (−) subjects (123 vs. 229).
Our third major finding was that the prediction model, which consists of age, gender, APOE genotype, and global SUVR, showed good performance (AUROC = 0.814). Note that despite being in the subthreshold range, the inclusion of global SUVR increased the performance. A few recent studies on longitudinal amyloid PET imaging have shown that the annual change rate of Aβ is biphasic [6, 31, 32]. The deflection point of this biphasic curve is known to be higher than the Aβ threshold. In subjects with baseline SUVR lower than this deflection point, the rate of Aβ accumulation increases as the baseline SUVR increases. In line with this pattern, our findings showed that the relative proximity of global SUVR to the Aβ threshold is a crucial factor in the classification of Aβ (−) subjects into converters and non-converters. External validation with the model using demographic features and global SUVR showed excellent performance (AUROC = 0.900), which means that the developed model can be useful in the clinic by discerning candidates who might convert to Aβ (+).
Moreover, the addition of regional SUVR led to an improvement in the prediction performance (AUROC = 0.841). This result is consistent with findings from previous studies that suggest that individuals with focal Aβ accumulation and negative global SUVR demonstrated early clinical and neuroimaging features of AD progression [33, 34]. Of note, the increased specificity and decreased sensitivity indicate that the model becomes more stringent with the incorporation of regional SUVR. It can be inferred that the model filters out subjects with high baseline SUVR but less risky regional uptake patterns. However, this needs further validation in a dataset with regional uptake values available. The incorporation of regional SUVR values could provide a more detailed understanding of the relationship between regionally specific amyloid aggregation and amyloid-related neurodegenerative changes. In the present study, the combined model that used demographic features, global SUVR, and regional SUVR resulted in the best performance, suggesting a better capacity to predict the conversion of Aβ positivity compared to other models tested. To our knowledge, no studies have developed machine learning classifiers to predict amyloid conversion in Aβ (−) subjects.
This study identifies the list of highly influencing features: global SUVR and APOE ɛ4 carrier status contributed the most to the prediction of Aβ conversion, followed by regional SUVRs in the medial and lateral parietal, medial and lateral frontal, and cingulate cortices. Our results are in line with previous studies on the early involvement pattern of cortical Aβ accumulation [35]. Thus, our model is likely to capture early AD patterns of amyloid PET.
Overall, the model developed in this study was able to predict the conversion of Aβ positivity in Aβ (−) subjects. The model performance improved with the inclusion of global and regional SUVRs and achieved good performance, which was validated in an independent dataset.
Limitations
A few limitations of our study need to be noted. First, the sample sizes of the datasets used in our study were modest. In fact, the ADNI is the cohort with the largest number of longitudinal amyloid PET data. However, the SMC dataset had a smaller number of eligible subjects especially when limited to initially negative subjects who had longitudinal PET data. Despite the limitations of the sample size, we were able to validate our results. Accumulation of amyloid PET data, especially in amyloid-negative individuals, is needed for a more robust validation of our results. Second, while model 3 showed the best performance, the sensitivity and positive predictive value were relatively low. We can adjust the prediction score threshold of the neural network model to find a different balance between specificity and sensitivity depending on the purpose of the prediction model. In contrast to the results in the ADNI cohort, the specificity was relatively low in the SMC cohort. The difference may be attributed to the smaller number of subjects or differences between the cohorts, such as ethnicity, gender, educational attainment, and family history. Since the false positivity may pose ethical challenges in applying the model in clinical trials, the prediction score threshold of the model may have to be adjusted in favor of specificity rather than sensitivity. More importantly, incorporating additional features such as neuropsychological test results, other neuroimaging phenotypes, or genetic factors is needed to improve the overall performance of the model. Third, we could not test the model including regional SUVR with the SMC dataset, although the model showed the best performance in the ADNI cohort. SMC subjects were recruited from a PET registry comprising amyloid PET scans of three different tracers, which forced us to use the CL method for harmonization. Unfortunately, the application of the CL method for regional uptake has not yet been validated. Once the methodology regarding the regional application of the CL is validated, it needs to be tested. Fourth, we used CL values to validate a model developed using global SUVR values. While this was possible since CL values have a strong linear relationship with FBP SUVR [36], further validation studies with harmonized features are warranted. Nevertheless, it is noteworthy that we developed well-performing models for the prediction of Aβ conversion and found important features that should be considered for the selection of primary prevention of AD.