Early Detection of Alzheimer Disease Leveraging Artificial Intelligence: Where are We?

Welcome to Health Connect, the podcast for health professionals through which we will share the latest news and information on science and technology in the medical field. Today, we will talk about three relevant advances in artificial intelligence in the field of Alzheimer disease.

Did you know that, in people with Alzheimer disease, neuronal death combined with neurofibrillary tangles and amyloid plaques leads to cortico-cortical disconnections and infiltration of the locus ceruleus? This affects the temporoparietal association areas, causing the development of visual and attention disorders.¹

We know that approximately 47 million people are living with dementia worldwide. Its most common cause is Alzheimer disease, which contributes to up to 80 percent of cases and results in impairment of multiple cognitive domains. To date, preventing or slowing this decline is more promising than reversing it, and existing treatments that can improve symptoms are more effective in those patients who are in preclinical or very early stages of the disease.^1,2

Even years before the overt symptoms of Alzheimer disease appear, its pathology may be present in the patient, so both dementia and cognitive decline could go unnoticed.^{1, 2} However, the identification of high-risk patients remains a considerable challenge. A common approach is to monitor patients with signs of mild cognitive impairment,³ a condition that in some cases can be a prodromal stage of dementia, especially in Alzheimer disease, and whose main characteristic is minor memory impairment. Then again, this mild cognitive impairment is truly an early stage of Alzheimer only in a subgroup of patients, so it is important to differentiate them from those who will remain stable.²

This trend occurs even in specialized settings; for example, memory clinics have reported an annual conversion rate of patients with mild cognitive impairment of 9.6 percent, showing that, even after ten years, patients with mild symptoms did not progress to dementia.³

This is all the more compelling because the increasing life expectancy results in more people being diagnosed with neurocognitive disorders.² However, existing diagnostic tools can estimate the onset of clinical dementia only in a medium- and long-term time frame in specific populations.³

For this reason, there has been increasing interest in the development of diagnostic automation methods for the early detection of neurodegenerative diseases. Researchers Grueso and Viejo-Sobero recognize that the field is promising in their systematic review of studies that used machine learning methods to predict progression of Alzheimer disease over the past decade.²

The majority of the 116 studies that they analyzed used magnetic resonance imaging, achieving a mean accuracy of 74.5 percent, as well as positron emission tomography data, which achieved a mean accuracy of 76.9 percent. The combination of both techniques reached even better results, with an average accuracy of 77.5 percent.²

A less frequent method, magnetoencephalography, reported an accuracy that ranged from 87 percent to 100 percent, but the sample sizes were very limited, being the smallest of 33 subjects and the largest of 129. In comparison, the mean sample size of all the studies reviewed was 546 participants.²

Among the most relevant selected features were whole-brain volumes, intensity measurements of glucose metabolism, genetic features, neuropsychological test results, and demographic variables such as age. The most useful brain areas to distinguish patients with Alzheimer, or stable mild cognitive impairment, from healthy subjects were mainly located in the temporal, parietal, and frontal lobes, while the relevant regions were the hippocampus, amygdala, entorhinal cortex, precuneus, cingulate gyrus, and rostral and caudal areas of the medial frontal lobe.²

The machine learning methods most used to classify the patients and detect disease progression were based on support vector machines, which reached a mean accuracy of 75.4 percent, although convolutional neural networks attained a better result with 78.5 percent. It is important to highlight that more complex models that combined deep learning with multimodal and multidimensional data achieved the best performance.²

Even so, the researchers warn that although the algorithms are useful and capable of discriminating the brain characteristics of Alzheimer disease, their performance is not specific enough. Thus, the judgment by a health professional remains crucial, because clinical criteria can still achieve similar results regarding the prediction of conversion to Alzheimer dementia up to a year before its onset.²

A frequent limitation in machine learning models is that they often incorporate complex variables, such as neuroimaging data, genetic tests, and cerebrospinal fluid biomarkers that are hardly available in routine clinical practice.³

In this context, let us explore the findings of the study published by James and colleagues in December 2021, where they evaluated the ability of machine learning algorithms to predict the incidence of dementia within two years. Importantly, the authors determined which variables were required to achieve an optimal analytical approach.³

In this prognostic study, the researchers studied a prospective cohort of approximately 15 300 patients from a memory clinic who did not have dementia at baseline. The median patients’ age was 72.3 years. Within two years of onset, ten percent of the participants received a diagnosis of dementia, divided by subtype as follows: 1 285 had Alzheimer dementia; 82 had Lewy body dementia; 21 had vascular dementia, and 180 had other subtypes of dementia.³

With these data, four machine learning algorithms were implemented to perform a classification task based on baseline variables that were recorded at the participants' first visit to the memory clinic. The models used were logistic regression, support vector machine, random forest, and gradient-boosted trees. The 258 variables that were used as candidate predictors encompass domains of clinical measures and risk factors related to dementia.³

Next, all machine learning models were evaluated by comparing their performance with each other, and with that of two existing tools for dementia risk prediction. The first was the Cardiovascular Risk Factors, Aging, and Incidence of Dementia risk score, abbreviated as CAIDE, which is used to determine the 20-year risk of developing dementia in middle-aged people. The second tool was the Dementia Screening Brief Indicator, or BDSI, which calculates the risk that elderly patients have in a six-year projection. Comparing these tools with each other, the BDSI performed better than the CAIDE, probably because it was designed specifically for older adults and its six-year follow-up period is more moderate.³

Nevertheless, all machine learning models achieved a better prediction of dementia development at two years compared with both tools. When comparing among models, they all performed similarly, with the gradient-boosted trees algorithm achieving the best results using all 258 variables, with a mean overall accuracy of 92 percent, a sensitivity of 0.45, a specificity of 0.97, and an area under the curve of 0.92.³

Then, when evaluating the performance of each model in differentiating dementia subtypes, logistic regression was the best at identifying dementia. This model correctly classified 46 percent of participants with Alzheimer dementia and 55 percent with other dementia subtypes. The support vector machine performed better in identifying Lewy body dementia, correctly classifying 49 percent of participants. All models correctly classified participants with vascular dementia.³

Another great drawback of using the machine learning approach is the large number of variables that are required to obtain a good performance, because as the number of required variables increases, the clinical implementation becomes less practical and the interpretability is reduced. James and researchers addressed this problem by evaluating the variability of the area under the curve according to the number of variables included.³

Although the models required 22 of the original 258 variables to achieve a diagnostic performance indistinguishable from their optimal mean performance, six variables used in all models were highly predictive. These variables were the clinical judgment of decline, the time to complete the Trail Making Test Part B for dementia, the level of independence, and three components of the Clinical Dementia Rating scale, specifically orientation, memory, and impairment of home and hobbies.³

Training each model using only these six variables, both the logistic regression and gradient-boosted trees models reached an accuracy of 91 percent and an area under the curve of 0.89, so there was no significant decrease in diagnostic performance.³

Finally, one of the most relevant findings is that the study identified that eight percent of patients labeled as having dementia appeared to be misdiagnosed, as the diagnosis was later reversed. The random forest model accurately identified 84 percent of these inconsistent diagnoses, while linear regression had the lowest percentage at 70.8 percent. This could prove that machine learning models can correctly identify patients who may have been misdiagnosed. Therefore, they could be the basis for developing a decision-making aid tool and clinical validation in specialized clinical settings.³

CAPSULE

It has been reported that most studies that achieve higher levels of precision in differentiating patients with Alzheimer from healthy controls tend to use specialized data combined with increasingly complex classification methods.¹

Despite this, we need to think about how adaptable these algorithms can be in daily clinical practice. So, it is important for new research to focus on achieving high performance using only clinical data commonly available to clinicians, such as structural magnetic resonance imaging, as well as demographic and cognitive measures. Future research should also make their models more generalizable, and with more diverse and inclusive samples.¹

END OF THE CAPSULE

Welcome back. In the previous segment, we explored the state of research on different machine learning approaches to predict the progression to Alzheimer disease. We will now talk about the possibilities of using machine learning combined with two preclinical signs that have been frequently reported in patients with Alzheimer, so they could help differentiate these patients from healthy people.

Several studies have reported that changes in eye movements and language are signs of cognitive decline at the onset of Alzheimer dementia. Even more important, the decline of both signs seems to progress along with the worsening of the disease.¹

Let's consider how these signs have been used in research. Language characteristics, such as semantic and lexical content and syntactic complexity, have been found to decline with the progression of the disease. Even speech patterns have shown an association with Alzheimer's pathology. In 2016, Fraser and colleagues managed to identify people with Alzheimer from healthy patients using only linguistic and acoustic characteristics of the speech, with an accuracy of 81.96 percent.¹

The ocular function is also altered by Alzheimer disease, so it is common to find abnormal saccadic behaviors and saccadic gaze intrusions, as well as slow pupillary responses and disorders. These abnormalities can be easily detected in tasks such as reading, as patients have slower reading and higher fixations. They reread words more frequently and are less likely to skip small, irrelevant words that a healthy reader might ignore.¹

Even more interesting is that language and eye movements seem to act synergistically. This is shown in recent work by Hyeju Jang and colleagues. Published in September 2021, it presents a machine learning analysis of a new multimodal eye-tracking and language dataset that can classify people with established Alzheimer disease, mild cognitive impairment, and subjective memory complaints from healthy controls.¹

It is worth mentioning that this work is based on the conference paper that the group published in 2020. That study was conducted with 68 patients at different stages of cognitive decline and 73 controls. The team used a description task known as "the cookie theft picture", in which patients are asked to describe an image depicting a domestic scene. The best performance using language-only models was an area under the ROC curve of 0.73, and of 0.77 for eye movement-only models. Later, when combining both eye movement and language models, the overall performance increased to 0.80.¹

From those results, Jang and team sought to more accurately classify patients with clinically significant cognitive impairment. The employed cohort is now larger, with 79 patients from a memory clinic diagnosed with mild-moderate Alzheimer, mild cognitive impairment, or subjective memory complaints, and 83 adult controls.¹

For this new study, the researchers incorporated two tasks. These were the description of a picture and the reading of a paragraph. The researchers added two new tasks, the calibration of the pupil, in order to capture potential square wave jerks that often occur in Alzheimer, and the description of a pleasant memory, to detect spontaneous speech data and deficiencies that may be overlooked in the established tasks.¹

Because the overall goal of the study was to build a screening tool rather than a diagnostic tool, all dementia subgroups were categorized as "patients" and compared against the healthy control group to identify highly predictive features shared across the entire disease spectrum.¹

After standardizing the data, a full set of features was extracted from each task to build task-independent models, testing each one against three different classification algorithms: logistic regression, random forest, and Gaussian naïve Bayes. These were chosen based on the results of the previous study, as they generated the best performances using both the eye-tracking and the image description task.¹

The findings proved that the novel tasks alone significantly outperform a dummy model, demonstrating that they have the capacity to discriminate between patients with clinically significant cognitive impairment versus controls. The pupil calibration achieved a classification accuracy of 71 percent, while the memory description had a better result of 78 percent. When comparing the four tasks, the established ones outperformed the novel ones, validating their use for classification in future research.¹

The next step was to determine the synergy between the tasks. To do this, the researchers developed a fusion model that combines the predictions of the four tasks. Fusion models based on logistic regression significantly outperformed all individual task models. All other fusion models achieved better results than most of the individual ones, obtaining particularly improved performance in the case of image description and reading tasks. This suggests that task fusion has a synergistic effect, increasing model performance.¹

Lastly, the researchers looked at how important each feature is for performance. The relevant findings reveal that in the pupil calibration task, the patients showed more variation and more eye movements, so they may be more prone to refixation.¹

On the other hand, acoustic features were more important than other language features. Speech characteristics, such as pause and speed, captured by acoustic analysis could have a high discriminatory capacity, since language performance was similar between reading tasks, where vocabulary is limited, as well as memory and description tasks, where spontaneous speech is analyzed.¹

Similarly, pupil calibration had an area under the curve of 0.71, consistent with the other tasks involving eye movements. The pupil calibration task lasts for ten seconds, where the participant looks at a fixed point on the screen while an infrared eye tracker records gaze data and pupil size.¹

Considering its simplicity and speed, pupil calibration could be a very good candidate for high-throughput detection. In addition, the results suggest that this task may be able to capture abnormal saccadic behavior that can be attributed to Alzheimer-related amyloid plaques in the brainstem, which affect premotor neurons responsible for generating saccades.¹

Finally, it is interesting to consider that this screening model seems to be highly tolerable. The application of the four tests lasted ten minutes, and more than 90 percent of the participants stated that they felt comfortable and interested during the evaluation. The majority were willing to repeat the evaluation in a clinical setting on an annual basis, although they felt less inclined to do it more often. Only 11 percent reported discomfort during testing, and five percent reported privacy issues with technology.¹

These figures are especially relevant if we take into account that, although the initial screening is important, the follow-up screening is also key for detecting longitudinal changes as the disease progresses.¹ The reason for this is that there is a critical period in which the morphological and functional changes in the central nervous system allow timely initiation of clinical treatment, which could delay the development of the disease.² Therefore, it is necessary to have non-invasive and comfortable tools that allow participants to agree to repeat the screening periodically.¹

Thanks for joining us on this episode of Health Connect. Don't miss our next episode, where we will talk about more artificial intelligence developments in other medical fields. Discover more medical news and content on Viatris Connect.

References:

1. 1. Jang H, Soroski T, Rizzo M, Barral O, Harisinghani A, Newton-Mason S, et al. Classification of Alzheimer's Disease Leveraging Multi-task Machine Learning Analysis of Speech and Eye-Movement Data. Front Hum Neurosci. [Internet] 2021. (Accessed on June 11, 2022);15:716670. Available at: https://www.frontiersin.org/articles/10.3389/fnhum.2021.716670/full
2. 2. Grueso S, Viejo-Sobera R. Machine learning methods for predicting progression from mild cognitive impairment to Alzheimer's disease dementia: a systematic review. Alzheimers Res Ther. [Internet] 2021. (Accessed on June 11, 2022);13(1):162. Available at: https://alzres.biomedcentral.com/articles/10.1186/s13195-021-00900-w
3. 3. James C, Ranson JM, Everson R, Llewellyn DJ. Performance of Machine Learning Algorithms for Predicting Progression to Dementia in Memory Clinic Patients. JAMA Netw Open. [Internet] 2021. (Accessed on June 11, 2022);4(12):e2136553. Available at: https://jamanetwork.com/journals/jamanetworkopen/fullarticle/2787228

Os links para todos os sites de terceiros são oferecidos como um serviço aos nossos visitantes e não implicam endosso, indicação ou recomendação do Health Connect. Os artigos vinculados são fornecidos apenas para fins informativos e não visam implicar uma atribuição pelo autor e/ou editor. O Health Connect se isenta de qualquer responsabilidade pelo conteúdo ou pelos serviços de outros sites. Recomendamos que você analise as políticas e condições de todos os sites que escolher acessar.

NON-2022-13045
NON-2023-2512