Prediction of neuropathologic lesions from clinical data

Abstract

Introduction: Post-mortem analysis provides definitive diagnoses of neurodegenerative diseases; however, only a few can be diagnosed during life.

Methods: This study employed statistical tools and machine learning to predict 17 neuropathologic lesions from a cohort of 6518 individuals using 381 clinical features (Table S1). The multisite data allowed validation of the model's robustness by splitting train/test sets by clinical sites. A similar study was performed for predicting Alzheimer's disease (AD) neuropathologic change without specific comorbidities.

Results: Prediction results show high performance for certain lesions that match or exceed that of research annotation. Neurodegenerative comorbidities in addition to AD neuropathologic change resulted in compounded, but disproportionate, effects across cognitive domains as the comorbidity number increased.

Discussion: Certain clinical features could be strongly associated with multiple neurodegenerative diseases, others were lesion-specific, and some were divergent between lesions. Our approach could benefit clinical research, and genetic and biomarker research by enriching cohorts for desired lesions.

Keywords: Alzheimer's disease neuropathologic change; Lewy body disease; comorbidities; neuropsychological battery tests; post-mortem autopsies.

FIGURE 1.

Overall prediction process and the counts and correlations within clinical data and neuropathologic lesions. (A) Schematic diagram of the process starting from cohort (with description shown) to train/test splitting for cross-site validation. The subfigure also shows a schematic diagram of the multitask LSTM model that takes longitudinal clinical features as input and outputs a prediction of 17 neuropathologic changes for each individual. The 381 clinical features (Figure 1B, full list shown in Table S1) of the cases in the training set (35 sites out of the 37 sites) were fed to the model to train to predict 17 neuropathologic lesions (Figure 1C). Note that, the correlation network (Figure 1B, see Supplementary Methods for details), where each node represents a feature and each edge signifies a strong correlation between the two, suggests strong correlation among clinical features within the same modality as seen by the proximity of their nodes. The model performance was then evaluated on the two unseen test sites. This cross-site validation process was repeated 5 times to obtain the average prediction performance. More details can be found in the method section. (B) A correlation network with a t-Distributed stochastic neighbor embedding (t-SNE) derived layout showing correlation of clinical features where each node is a clinical feature and edges represent the top 10% of Spearman’s correlation P-value among feature pairs. t-SNE is a nonlinear dimension reduction algorithm that took in all 381 clinical features as input and reduced the information into two dimensions, where similar features were put closer together, to allow feature visualization. 65 The node is colored by the modality of the features. Each modality is described along with the number of features within it. More correlated clinical features situate closer together. (C) The correlations among neuropathologic lesions indicate a strong correlation (Spearman’s) among pTDP-43 inclusions in multiple brain regions and a moderate correlation of pTDP-43 inclusions (TDPs) to HIPSCL-SCL, ARTE, and WMR. (D) The count and availability of each neuropathologic lesion in the NACC dataset. (E) The distribution of cognitive diagnosis at the last visit: dementia (Dem.) mild cognitive impairment (MCI), and no cognitive impairment (NCI) for each lesion. (F) A Venn diagram displaying the number of cases with ADNC, LEWY, and AMY comorbidities.

FIGURE 2.

Cross-site multitask model prediction indicates high performance for certain neuropathologic lesions and revealed a critical set of clinical features. (A) The cross-site LSTM model prediction performance was reported as the area under the receiver operating curve (AUROC) and area under the precision-recall curve (AUPRC) for all of the available lesions. Note that a random prediction would approximately yield an AUROC of 0.5 and AUPRC equivalent to the prevalence of that lesion (shown as gray patterned bars). (B) The ROC and PRC of the selected lesions, including ADNC, AVAS, LEWY, and TDPs. For comparison purposes in ADNC and LEWY, the “x” marks pinpoint sensitivity, specificity, and precision based on clinical diagnosis of AD dementia or dementia with Lewy body (DLB). (C) Model reduction by using only top-k clinical features and their resulting performance for the selected neuropathologic lesions (ADNC, AVAS, LEWY, and TDPs) to yield similar performance as using all clinical features, that is, the point where the AUROC started to plateau. (D) A heatmap of all top 70 clinical features for the lesions grouped into 13 different sections according to the cutoff on the hierarchical clustering shown in the red line.

FIGURE 3.

Example of strong, shared, or diverging clinical features for neuropathologic lesions. (A) Apart from clinician diagnosis, examples of other uniquely strong ADNC-correlated features are the presence of APOE ε4 allele, absence of bvFTD diagnosis, impaired memory as a predominant symptom, and cognition as the first area of change. The numbers on bar graphs indicate fold count ratios between with or without lesions. (B) Examples of features are moderately correlated among all features except for AVAS. (C) FTLD hereditary mutation is the strongly correlated feature for the TDP-Neocortex. (D) The opposite trend of ADNC (without TDP-neocortex) vs TDP-neocortex (also shown in section 10 of Figure 3B), such as personality change, where the presence of personality change is negatively correlated to ADNC, but positively for TDPs. (E) Apart from clinician diagnosis of DLB, hallucination is a uniquely correlated feature to LEWY. (F) Strongly correlated features of only AVAS are hypertension and age.

FIGURE 4

Higher number of comorbidities worsens cognitive test performance and some combinations of comorbidities with ADNC result in significantly different degrees of change in symptoms. (A) Fold change in neuropsychological battery test performance and clinical dementia rating (CDR) as the number of comorbidity increases. The fold change is calculated from the difference in the mean score of the selected combination of lesions and the mean score of ADNC^low divided by the mean score of ADNC^low. Each dot represents a test or rating type. The comparison was controlled for age by only selecting cases within ± 1SD of the mean age of cases with only one neuropathologic lesion. (B) The fold change in ADNC^low with different comorbidity combinations (with at least 15 cases) compared to cases with pure ADNC^low (n = 42) of the same age range. The size of each bubble indicates the Q value from the Mann-Whitney U test, whereas the color intensity represents the fold change min-max normalized within each test type. The fold change number is written in each bubble with negative sign inverted, that is, positive numbers indicate worse performance. (C) An example of a significant change or no change in trail-making test B and CDR–Memory of ADNC cases with selected comorbidities compared to pure ADNC^low.

FIGURE 5

Multitask LSTM model can improve ADNC cohort purity by predicting ADNC without selected comorbidities. (A) The cross-site LSTM model prediction AUROC and AUPRC for the prediction of clinical AD cases without each of the comorbidities (positive class). Only cases with clinically diagnosed AD dementia were considered. The patterned bars indicate the prevalence of AD dementia with each neuropathologic comorbidity. (B) The ROC and PRC of the interested comorbid lesions, including LEWY, HIPSCL-SCL, and TDPs. For comparison in LEWY comorbidity, the “x” marks pinpoint sensitivity, specificity, and precision based on LBD, the only clinically available diagnosis. (C) Top univariate features that best separate clinical AD cases without TDP-Amygdala or TDP-Neocortex comorbidity. (D) Cohort size versus positive predictive value (PPV) curve to visualize the performance of the model for purifying ADNC-focused cohorts.