AI May Detect Alzheimer's Years Before Diagnosis

Batya Swift Yasgur, MA, LSW

November 06, 2018

An artificial intelligence (AI) algorithm that uses positron-emission tomography (PET) improves the ability of brain imaging to predict Alzheimer's disease (AD) at an early stage, new research shows.

Investigators studied more than 2000 prospective 18F-fluorodeoxyglucose (18F-FDG) PET images taken from 1000 patients from the Alzheimer's Disease Neuroimaging Initiative (ADNI). They trained the algorithm on 90% of the dataset and then tested it on the remaining 10%.

The algorithm successfully "learned" to identify the metabolic patterns that corresponded to AD.

When the algorithm was tested on an independent set of 40 images from 40 never-studied patients, it achieved 100% sensitivity at detecting the disease an average of more than 6 years prior to the final diagnosis.

"The key point of our study is that our algorithm not only successfully detects AD but actually detects it 6 years before the diagnosis is made," corresponding author Jae Ho Sohn, MD, of the Department of Radiology and Biomedical Imaging, University of California, San Francisco, told Medscape Medical News.

"In order to develop treatments for AD, and even for the patient's sake, it is better to know about the disease early, because by the time it is diagnosed, there is usually too much brain volume loss and little we can do to help the patient," he said.

The study was published online November 6 in Radiology.

Precision Algorithm

Advances in diagnostic technology, such as 18F-FDG PET imaging, allow earlier diagnosis and treatment of AD, but 18F-FDG PET currently requires interpretation by nuclear medicine neuroimaging specialists "to make pattern recognition decisions mostly using qualitative readings, [which is] particularly challenging in the setting of a disease that involves a wide continuous spectrum from normal cognition to MCI [mild cognitive impairment] to AD," the authors write.

"Deep learning" may "assist in addressing the increasing complexity and volume of imaging data, as well as the varying expertise of trained imaging physicians," they add.

Although this model has been studied with respect to other illnesses, its application to brain imaging is "only beginning to be explored," the authors note.

"There has long been suspicion that some pattern in the way we study FDG uptake in the brain can signify signs of AD or be an early prediction of AD developing over time, but unfortunately, there has not been a really definitive method," Sohn remarked.

"Unlike a brain tumor or seizure process, where you can actually see a focal area lighting up, AD development is subtle, and it is diffuse and present throughout the brain — and although it shows some predilection for different regions, there are no focal findings," he explained.

"PET artificial intelligence is gaining immense popularity in the press and research as a type of precision algorithm that is nonlinear in property and allows us to recapture subtle yet diffuse findings in an efficient manner," he said.

To investigate whether a deep learning algorithm (deep learning model Inception V3) could be trained to predict the final clinical diagnoses of patients who underwent PET brain imaging and, once trained, how its results compare to the actual diagnoses arrived at by current clinical reading methods, the researchers employed 2109 prospective 18F-FDG PET brain images from ADNI imaging studies conducted from 2005 to 2017 (n = 1002 patients).

Of this dataset, 90% (1921 imaging studies, 899 patients) were used for model training and internal validation. The remaining 10% (188 imaging studies, 103 patients) were used for model testing; these images served as the internal test set.

The researchers also used an additional test set obtained from their institution (the "independent test set"). This set, which served as the external test set, comprised 40 18F-FDG PET imaging studies from 40 patients who were not enrolled in the ADNI. The dates for these imaging studies ranged from 2006 to 2016.

Final clinical diagnosis determined after all follow-up examinations was used as the ground truth label for both datasets.

Three nuclear medicine physicians independently interpreted the 40 18F-FDG PET imaging studies in the independent test set.

"Perfect" Sensitivity, "Reasonable" Specificity

The average age of the male patients in the ADNI study was 76 years (range, 55 years); the average age of the female patients was 75 years (range, 55 to 96) (P < .001). Overall, 54% of the patients were men (547 of 1002); by imaging study, 58% of the patients were men (1225 of 2109).

The average follow-up period was 54 months by patient and 62 months by imaging study.

Of the 40 patients in the independent test set, seven were clinically diagnosed as having AD, seven as having MCI, and 26 as having non-AD/MCI at the end of the follow-up period.

The average age of these male patients was 66 years (range, 48 to 84 years); for female patients, the average age was 71 years (range, 41 to 84).

The overall percentage of men in the independent test set was 58% (23 of 40). The average follow-up period of patients in the independent test set was 76 months. In the AD group, the average follow-up was 82 months; in the MCI group, it was 75 months; and in the non-AD/MCI group, it was 74 months.

Inception V3 was trained on 90% of ADNI data and was tested on the remaining 10%. The receiver operating characteristic (ROC) curve of the deep learning mode yielded an area under the curve (AUC) for prediction of AD of 0.92; for MCI, it was 0.63; and for non-AD/ MCI, it was 0.73.

These findings "indicated that the deep learning network had reasonable ability to distinguish patients who finally progressed to AD at the time of imaging from those who stayed to have MCI or non-AD/MCI, but was weaker at discriminating patients with MCI from the others," the authors state.

The sensitivity for the prediction of AD, MCI, and non-AD/MCI was 81% (29 of 36), 54% (43 of 79), and 59% (43 of 73), respectively.

Specificity was also high, at 94% (143 of 152), 68% (74 of 109), and 75% (86 of 115), respectively.

Precision was 76% (29 of 38), 55% (43 of 78), and 60% (43 of 72), respectively.

The ROC tested on independent test set yielded an AUC for the prediction of AD, MCI, and non-AD/MCI of 0.98 (95% confidence interval [CI], 0.94 - 1.00), 0.52 (95% CI, 0.34 - 0.71), and 0.84 (95% CI, 0.70 - 0.99), respectively.

When the researchers chose the class with the highest probability as the classification result, the sensitivity was 100% (7 of 7), 43% (3 of 7), and 35% (9 of 26) for the prediction of AD, MCI, and non-AD/MCI, respectively.

The specificity was found to be 82% (27 of 33), 58% (19 of 33), and 93% (13 of 14), and the precision was 54% (7 of 13), 18% (3 of 17), and 90% (9 of 10) in the prediction of AD, MCI, and non-AD/MCI, respectively.

"With a perfect sensitivity rate and reasonable specificity on AD, the model preserves a strong ability to predict the final diagnoses prior to the full follow-up period that, on average, concluded 76 months later," the authors comment.

Compared with the radiology readers, the deep learning model performed statistically significantly better in recognizing patients who would go on to receive a clinical diagnosis of AD.

It also performed better on the independent test set at recognizing patients with neither AD nor MCI. However, it performed worse at recognizing patients who would develop MCI but whose condition would not progress to AD, although this finding was without statistical significance.

"In predicting the final diagnosis of AD on the independent test set, it outperformed three radiology readers in ROC space," the authors note.

"Although there were false positives, the fact that the algorithm could detect every case of AD is a big feat," said Sohn.

"I see this algorithm as complementing the work of radiologists, especially in conjunction with other biochemical and imaging tests," he added.

Increased Sophistication

Commenting on the study for Medscape Medical News, Arthur Toga, PhD, Laboratory of Neuroimaging, USC Stevens Neuroimaging and Informatics Institute, Keck School of Medicine of University of Southern California, Los Angeles, who was not involved with the study, said that the "authors trained a deep learning model that gave a more accurate prediction of AD and MCI than professional human radiology readers."

The authors also "provided the structure and hyper-parameters of their neural network model, which can be used as a reference for further improvement," he noted.

The findings have implications for clinical usage, Toga said. "As the sophistication of deep learning models continues to improve, we are certain to see wider adoption in clinical practice as a decision support tool."

He noted that although 18F-FDG PET is "one of the tools used in AD diagnosis," the high costs of scans, which the authors note, "remain a challenge."

Sohn added, "One limitation of our study is that it's on the smaller side, only 40 patients, so it calls for further validation with larger datasets at different institutions, which is a necessary step before the findings can be integrated into clinical care."

An accompanying editorial by Mykol Larvie, MD, of the Division of Neuroradiology, Department of Nuclear Medicine, Cleveland Clinic, Ohio, stated that the researchers' "machine learning application and description of the trial data" are "adequate for other researchers to replicate their analysis."

Data collection and sharing for the project was funded by the ADNI, the National Institutes of Health, and the US Department of Defense. Dr Sohn, Dr Larvie, and Dr Curfman have disclosed no relevant financial relationships. Coauthors' disclosures of such relationships are listed in the original articles.

Radiology. Published online November 6, 2018. Full text, Editorial

For more Medscape Psychiatry news, join us on Facebook and Twitter.

Comments

3090D553-9492-4563-8681-AD288FA52ACE
Comments on Medscape are moderated and should be professional in tone and on topic. You must declare any conflicts of interest related to your comments and responses. Please see our Commenting Guide for further information. We reserve the right to remove posts at our sole discretion.
Post as:

processing....