3517

Brain atrophy and machine learning algorithms on the prediction of dementia development
Pedro Henrique Rodrigues da Silva1, Kaio Felippe Secchinato1, Julia Palaretti1, and Renata Ferranti Leoni1
1USP, Ribeirão Preto, Brazil

Synopsis

A challenging issue regarding the early diagnosis of the Alzheimer's disease (AD) is the selection of biomarkers. In this study, we aimed to classify cognitively normal elderly regarding the possibility to develop AD based on brain atrophy and neuropsychological scores, and using supervised machine learning algorithms. Our results suggest Naïve-Bayes (NB) classifiers with left postcentral and left middle temporal cortical thickness or right lateral ventricle, right inferior parietal and Corpus Callosum (CC) Mid Posterior volumes can be useful to identify in the early stage the subjects with higher risks to develop AD.

Introduction

The early detection of Alzheimer's disease (AD) increases the benefits of treatments1. However, AD starts its development years before cognitive decline2. Therefore, early detection and diagnosis are crucial3. The assessment of cerebral atrophy is considered one important biomarker in the early stage of AD3. Machine learning algorithms have been widely applied in the diagnosis of AD patients4–9. Then, this study aimed to classify healthy elderly regarding the development of dementia using brain volume and cortical thickness measures, neuropsychological (NP) assessment scores and supervised machine learning algorithms. We hypothesized that regional atrophy and machine learning classifiers can be useful to indicate individuals with a higher risk of AD development.

Materials and Methods

MRI data of 54 subjects weRE obtained from the publicly available Open Access Series of Imaging Studies (OASIS) - OASIS310. Clinical Dementia Rating11 (CDR) scale was used to define the groups: twenty-seven cognitively normal subjects (CDR=0) who progressed to AD (CDR=1) after MRI acquisition (Converters, age: 79.9±6.4 years, 13 males), and twenty-seven cognitively normal subjects who remained stable (Stables, age: 79.7±6.0 years, 13 males). Scans were obtained on BioGraph mMR PET-MR 3T by the Knight Alzheimer Research Imaging Program (Washington University, St. Louis). T1-weighted images were acquired with the following parameters: slice thickness = 1.2 mm, TE = 0.00295 s, TR = 2.3 s, inversion time = 0.9 s, flip angle = 9°. FreeSurfer12 was used for cortical reconstruction and volumetric segmentation. Neuropsychological (NP) scores of the Mini-Mental State Examination (MMSE), Digit Span, Category Verbal Fluency (animals and vegetables), Trail Making Test (TMT) (TRAIL-A, TRAIL-B), Logical Memory - Story A, WAIS and Boston were also considered. Presence of other diseases and risk-factors were considered as confounders. Shapiro-Wilk Normality Test and Group Comparison were performed. We performed the feature selection using the Boruta method13 with Boruta package14 for R Software15 to reduce input dimensionality and avoid redundant or irrelevant features to the predictive model. The collinearity between the selected features was also assessed using Spearman's correlation16 since it can decrease the predictive capacity of regression or classification models. We considered the following input parameters: volume measures, volume and NP measures, thickness measures, thickness and NP measures, volume and thickness measures, and volume, thickness, and NP measures. Seven classifiers were trained and tested on the selected features R language (3.3.2) using caret17: Random Forest (RF)18, Linear and Radial Support Vector Machine (SVM)19, Naïve Bayes (NB)20, K-nearest neighbor (kNN)21 and an Artificial Neural Network (ANN)22. Data were divided into sets for training and testing the classifier using the k-folding cross-validation method23. We used 5-fold cross-validation, splitting the dataset into 70% training set and 30% testing set. Receiver Operating Characteristic (ROC) was used to select the optimal model using the highest value, except for the RF algorithm. Results were analyzed using confusion matrix indicators to assess the classifier's performance, including accuracy, precision, sensitivity , and specificity.

Results

Groups presented significant differences in MMSE (p = 0.028), and verbal fluency – category vegetables (p = 0.022), and a trend for TRAIL-A (p = 0.091). Traumatic brain injury, hypertension, hypercholesterolemia, diabetes, vitamin B12 deficits, thyroid disease, depression, apolipoprotein E4 (ApoEɛ4), alcohol abuse and smoking were not significantly different among groups (p>0.05). Table 1 shows the selected features. Figure 1 shows the performance of the classifiers using the training set. They achieved similar results by ROC values across different feature selection. Table 2 shows the performance of the classifiers using the testing set. KNN, NB, and RF presented similar performance with volumetric measures. When added NP measures, it improved the performance of the NB classifier, which achieves accuracy, sensibility, and sensitivity with the test set of 87%, 75%, and 100%, respectively. NB classifiers achieved the highest accuracy (87%) with cortical thickness measures, although the adding of NP variables collapses its accuracy to only 50%, making the KNN and Linear SVM the best classifiers. KNN and NB classifiers performed better with the volume and cortical thickness measures, although the Boruta feature selection and Spearman’s correlation methods produced only volumetric metrics as final features. When mixing volume, cortical thickness, and NP measures, KNN, Linear SVM, RF, and ANN achieved the best and similar performance. In general, the NB classifier with thickness measures and volume plus thickness measures produced the best accuracy and AUC.

Discussion

In general, NB classifier with thickness or volume measures only achieved the best accuracy from all the combinations in this study. Previously, in a multistage classification approach, the addition of NB improved the performance of the classification of AD patients in different stages of the disease7. Regarding the selected features, the left postcentral gyrus and middle temporal cortical thickness were significantly lower in AD patients than healthy controls24. Moreover, nitrated brain proteins in the inferior parietal lobule25, ventricular enlargements26, and callosal circularity27–29 have been associated with the early stage of AD.

Conclusion

Our results suggest that NB classifier with left postcentral gyrus and left middle temporal cortical thickness or right lateral ventricle, right inferior parietal lobule and CC volumes can identify cognitively normal subjects with higher risks to develop AD.

Acknowledgements

Data were provided by OASIS-3: Principal Investigators: T. Benzinger. D. Marcus. J. Morris; NIH P50AG00561. P30NS09857781. P01AG026276. P01AG003991. R01AG043434. UL1TR000448. R01EB009352.

This work was supported by the Fundação de Apoio à Pesquisa do Estado de São Paulo (FAPESP), process number 2017/22212-0

References

1. Pais, M. et al. Early diagnosis and treatment of Alzheimer’s disease: New definitions and challenges. Brazilian J. Psychiatry (2020) doi:10.1590/1516-4446-2019-0735.

2. Dubois, B. et al. Preclinical Alzheimer’s disease: Definition, natural history, and diagnostic criteria. Alzheimer’s and Dementia (2016) doi:10.1016/j.jalz.2016.02.002.

3. Nestor, P. J., Scheltens, P. & Hodges, J. R. Advances in the early detection of alzheimer’s disease. Nat. Rev. Neurosci. (2004) doi:10.1038/nrn1433.

4. Park, J. H. et al. Machine learning prediction of incidence of Alzheimer’s disease using large-scale administrative health data. npj Digit. Med. (2020) doi:10.1038/s41746-020-0256-0.

5. Nori, V. S. et al. Machine learning models to predict onset of dementia: A label learning approach. Alzheimer’s Dement. Transl. Res. Clin. Interv. (2019) doi:10.1016/j.trci.2019.10.006.

6. Moradi, E., Pepe, A., Gaser, C., Huttunen, H. & Tohka, J. Machine learning framework for early MRI-based Alzheimer’s conversion prediction in MCI subjects. Neuroimage (2015) doi:10.1016/j.neuroimage.2014.10.002.

7. Kruthika, K. R., Rajeswari & Maheshappa, H. D. Multistage classifier-based approach for Alzheimer’s disease prediction and retrieval. Informatics Med. Unlocked (2019) doi:10.1016/j.imu.2018.12.003.

8. Albright, J. Forecasting the progression of Alzheimer’s disease using neural networks and a novel preprocessing algorithm. Alzheimer’s Dement. Transl. Res. Clin. Interv. (2019) doi:10.1016/j.trci.2019.07.001.

9. Kim, H.-G. et al. Evaluation and Prediction of Early Alzheimer’s Disease Using a Machine Learning-based Optimized Combination-Feature Set on Gray Matter Volume and Quantitative Susceptibility Mapping. Curr. Alzheimer Res. (2020) doi:10.2174/1567205017666200624204427.

10. LaMontagne, P. J. et al. IC-P-164: OASIS-3: LONGITUDINAL NEUROIMAGING, CLINICAL, AND COGNITIVE DATASET FOR NORMAL AGING AND ALZHEIMER’S DISEASE. Alzheimer’s Dement. (2018) doi:10.1016/j.jalz.2018.06.2231.

11. Berg, L. Clinical Dementia Rating (CDR). Psychopharmacol. Bull. (1988).

12. Fischl, B. FreeSurfer. NeuroImage (2012) doi:10.1016/j.neuroimage.2012.01.021.

13. Kursa, M. B., Jankowski, A. & Rudnicki, W. R. Boruta - A system for feature selection. Fundam. Informaticae (2010) doi:10.3233/FI-2010-288.

14. Kursa, M. B. & Rudnicki, W. R. Feature selection with the boruta package. J. Stat. Softw. (2010) doi:10.18637/jss.v036.i11.

15. R Core Team (2019). R: A language and environment for statistical computing. Accessed 1st April 2019 (2019).

16. Zar, J. H. Spearman Rank Correlation. in Encyclopedia of Biostatistics (2005). doi:10.1002/0470011815.b2a15150.

17. Kuhn, M. & Contributions from Jed Wing, Steve Weston, Andre Williams, Chris Keefer, Allan Engelhardt, Tony Cooper, Zachary Mayer, Brenton Kenkel, the R Core Team, Michael Benesty, Reynald Lescarbeau, Andrew Ziem, Luca Scrucca, Yuan Tang, Can Candan, and T. H. caret: Classification and Regression Training. R Packag. version 6.0-79 (2018).

18. Breiman, L. Random forests. Mach. Learn. (2001) doi:10.1023/A:1010933404324.

19. Cortes, C. & Vapnik, V. Support-Vector Networks. Mach. Learn. (1995) doi:10.1023/A:1022627411411.

20. Set, A. Naive Bayes Classifiers. homepages.inf.ed.ac.uk (2009).

21. Peterson, L. K-nearest neighbor. Scholarpedia (2009) doi:10.4249/scholarpedia.1883.

22. Dayhoff, J. E. & DeLeo, J. M. Artificial neural networks: Opening the black box. in Cancer (2001). doi:10.1002/1097-0142(20010415)91:8+<1615::aid-cncr1175>3.0.co;2-l.

23. Refaeilzadeh, P., Tang, L. & Liu, H. Cross-Validation. in Encyclopedia of Database Systems (2009). doi:10.1007/978-0-387-39940-9_565.

24. Yang, H. et al. Study of brain morphology change in Alzheimer’s disease and amnestic mild cognitive impairment compared with normal controls. Gen. Psychiatry (2019) doi:10.1136/gpsych-2018-100005.

25. Reed, T. T., Pierce, W. M., Turner, D. M., Markesbery, W. R. & Allan Butterfield, D. Proteomic identification of nitrated brain proteins in early Alzheimer’s disease inferior parietal lobule. J. Cell. Mol. Med. (2009) doi:10.1111/j.1582-4934.2008.00478.x.

26. Nestor, S. M. et al. Ventricular enlargement as a possible measure of Alzheimer’s disease progression validated using the Alzheimer’s disease neuroimaging initiative database. Brain (2008) doi:10.1093/brain/awn146.

27. Ardekani, B. A., Bachman, A. H., Figarsky, K. & Sidtis, J. J. Corpus callosum shape changes in early Alzheimer’s disease: An MRI study using the OASIS brain database. Brain Struct. Funct. (2014) doi:10.1007/s00429-013-0503-0.

28. Van Schependom, J., Niemantsverdriet, E., Smeets, D. & Engelborghs, S. Callosal circularity as an early marker for Alzheimer’s disease. NeuroImage Clin. (2018) doi:10.1016/j.nicl.2018.05.018.

29. Bachman, A. H., Lee, S. H., Sidtis, J. J. & Ardekani, B. A. Corpus callosum shape and size changes in early Alzheimer’s disease: A longitudinal MRI study using the oasis brain database. J. Alzheimer’s Dis. (2014) doi:10.3233/JAD-131526.

Figures

Table 1: Final selected features obtained from Boruta’s method and exclusion of redundant variables using Spearman’s correlation.

Figure 1: Box plot comparing model results in the training set using the Caret R Package: (a) volume measures, (b) volume and NP measures, (c) thickness measures, (d) thickness and NP measures, (e) volume and thickness measures, (f) volume, thickness, and NP measures. Spec = Specificity. Sens = Sensibility. ROC = Receiver Operating Characteristic.

Table 2: Accuracies, sensibilities, specificities, Positive Predictive Value (PPV), Negative Predictive Value (NPV), and AUC (area under curve) calculated for each classifier in the testing set.

Proc. Intl. Soc. Mag. Reson. Med. 29 (2021)
3517