In this study, we developed and validated a prediction model based on a combination of clinical and radiomics features that could predict Aβ positivity based on CSF analysis at single subject level. The combined model involving both clinical and radiomic features showed the best performance (AUC: 0.823), followed by clinical model (AUC: 0.758), and the radiomics model (AUC: 0.674) in the test set, showing the utility and robustness of the combined model. These results indicate the independent contribution of radiomics and clinical features in identifying MCI with CSF Aβ pathology and the added value of the radiomics beyond the effects of clinical features.
Accumulation of Aβ pathology is one of the hallmark pathologic characteristics of the AD continuum and precedes decades before the onset of cognitive symptoms. (6) Recently, many amyloid-modifying therapy trials in AD subjects failed to show its effectiveness, (25–27) and one of the presumed reason for failure is the enrollment of subjects with clinical heterogeneity who did not have increased cerebral Aβ plaques and were unlikely to have had AD pathology. (28) Therefore, the identification of Aβ biomarkers via CSF Aβ or PET is important to diagnose the AD continuum in both research and clinical settings. However, these biomarkers are not routinely acquired in clinics, owing to limited resource, high costs, and the need for invasive procedures. Therefore, practical methods to determine candidates for the amyloid biomarker test with commonly available clinical and MRI data may be helpful.
Compared to MCI subjects without amyloid pathology, those with amyloid pathology have significantly lower volumes in various brain regions including the hippocampus. (29–32) Previous studies have already attempted to predict the amyloid pathology using these MRI features in MCI patients. They mainly used volume features of the hippocampus and other AD vulnerable structures, (29, 31, 33) offering a fair degree of diagnostic performance. Predictive models that combined both MRI and clinical features showed good performance in identifying amyloid pathology of MCI patients. (29–31, 33) Consistent findings were found in the current study. However, the most previous studies were performed without proper validation in a test set, which may have led to overfitted results, especially in high-dimensional datasets with machine learning studies. (34) A recent study applying data-driven algorithm with clinical features with validation showed an AUC of 0.71 at the test set, (35) showing only fair performance, unlike that in previous studies. This gives another line of evidence of potential overfitted results of the previous studies. Meanwhile, our model that integrated radiomics and clinical features showed good performance not only in the training set, but also in the test set. The robust predictive capacity of our combined models shows that it can help triage the subjects for more invasive and costly Aβ testing.
Although previous radiomics studies in the neuroradiology field have mostly been focused on neuro-oncology, (36–40) there have also been several recent studies using radiomics analysis on T1-weighted images in AD. These studies have shown promising results not only in the diagnosis of AD but also in the prediction of disease progression. (41–44) Radiomics features may be prone to biological validation for their correlation with disease pathology. (45) This observation is based on the hypothesis that radiomics features, especially second-order features, capture the spatial variation in signal intensity that may reflect the deposition of Aβ plaques. Further, it may extract different biological information from volume, (43, 46) which is the traditional imaging biomarker of AD.
Notably, nearly all the radiomics features were retained in the combined model after the LASSO procedure in our study. This suggests that most radiomics features harbor information independent from the clinical features, which may provide added value in predicting the CSF Aβ status. However, the prediction of CSF Aβ status by radiomics features alone was not optimal, confirming the importance of clinical features. Nonetheless, our results indicate that the added value of radiomics features over clinical features is a robust method.
Our study has several limitations. First, we only included the radiomics features of the hippocampus, as previous studies showed good performance using the hippocampus mask for the classification and prediction of AD. (40, 42, 47, 48) However, volume changes not only occur in the hippocampus, but also in other AD signature regions such as the entorhinal cortex and precuneus. (6) Thus, the radiomics prediction model could be improved by adding radiomic information of other anatomical structures. Further, whole brain investigation should be performed in future studies. Second, CSF Aβ status was used as the gold standard for Aβ positivity rather than PET imaging. It could be argued that the performance of the prediction model could be sensitive to the selection of the gold-standard method. However, the agreement between CSF and PET determinations of Aβ positivity is very high, particularly in the intermediate ranges where thresholds for positivity typically lie. (49, 50)