Leveraging Technology to Diagnose Alzheimer’s Disease: A Systematic Review and Meta-Analysis

Background: About 50 million people worldwide suffered from dementia in 2018—two-thirds of those with Alzheimer’s disease (AD). By 2050, this number is expected to rise to 152 million—which is slightly larger than the country of Russia. About 90% of these people are over the age of 65, but early-onset dementia can occur at younger ages. Early detection is imperative to expedient treatment, which can improve outcomes over the span of diagnosis. Objectives: To conduct a meta-analysis of similar studies along with a systematic literature review to hasten the development of clinical practice guidelines to assist clinicians in the diagnosis of AD. We analyzed data points in each article published over the last 10 years to meet this objective: cost, efficiency, accuracy, acceptability (by physician and patient), patient satisfaction, and barriers to adoption. Methods: Four research databases were queried (PubMed, CINAHL Ultimate, Web of Science, and ScienceDirect). The review was conducted in accordance with a published protocol, the Kruse Protocol, and reported in accordance with PRISMA (2020). Results: Ten interventions were identified to help diagnose AD among older patients, and some involved a combination of methods (such as MRI and PET). The average sample size was 320.32 (SD = 437.51). These 10 interventions were identified as accurate, non-invasive, non-stressful, inexpensive, convenient, and rapid. Only one intervention was identified as statistically ineffective, and this same intervention was used effectively in other studies. The barriers identified were cost, training, expense of travel, and required physical presence of patient. The weighted average sensitivity was 85.16%, specificity was 88.53, and the weighted average effect size was 0.7339 (medium). Conclusion: Innovation can accurately diagnose AD, but not all methods are successful. Providers must ensure they have the proper training and familiarity with these interventions to ensure accuracy in diagnosis. While the physical presence of the patient is often required, many interventions are non-invasive, non-stressful, and relatively inexpensive.


Introduction 1.Rationale
Prior to the pandemic, Alzheimer's disease (AD) was growing into the largest fear for older adults.In the US, AD is fatal for more people than breast cancer and prostate cancer, and it is the number one cause of death in Great Britain.The 2018 prevalence of dementia was 50 million, but it is expected to surpass 150 million by 2050.Alzheimer's disease accounts for about two-thirds of the dementia population [1].
The disease has no cure: there are only about 10 drugs approved to manage AD's symptoms.The disease manifests itself in the form of plaque on the brain that affects the communication of about 100 billion neurons, which degrades and ultimately destroys these neurons.At first, this destruction manifests itself in simple forgetfulness of recently learned things, but over time, the destruction becomes more severe, affecting speech, motor movement, and long-term memory [2].Medicine today does not fully understand the etiology and pathogenesis of AD [3].Medicine can treat but cannot prevent or cure the disease [4,5].For these reasons, many healthcare professionals have promoted research related to the detection of this disease because more timely diagnoses could potentially reduce the overall burden faced by society [2].
For a short period, diagnosis could only be confirmed through autopsy; however, health information technology (HIT) has made strides toward diagnosis while the patient lives.Optical and electrochemical biosensors are being developed to sense biological responses and identify biomarkers for AD [6].Canada recently published a white paper on the early diagnosis of AD.It included questionnaires and cognitive testing, diagnosis of neurodegeneration through MRI, and established biomarkers for neurodegeneration and Alzheimer's pathology through imaging and biomarkers [7].Wearables and other remote sensors track behavior and can help detect neurodegenerative diseases like AD in a timely and economical manner [8].
A systematic review was published in 2022 that evaluated apps, sensors, and virtual reality developed to help with timely diagnosis, management, and treatment of symptoms.Eight studies were analyzed from the last decade to identify solutions: deficits in finger dexterity, memory retrieval, and alertness and mood improvement [9].This study did not identify how leveraging health information technology can diagnose AD.
A systematic review was published in 2021 that analyzed health monitoring and artificial intelligence (AI) for deep learning.This review found AI can be used in the early detection of chronic diseases.Cloud computing was a catalyst for this innovation, and the integration of a blockchain framework improved data security to help prevent the misuse of patient data [10].This study did not identify how leveraging health information technology can diagnose AD.

Objectives
The purpose of this review is to conduct a meta-analysis of similar studies along with a systematic literature review to hasten the development of clinical practice guidelines to assist in the diagnosis of AD.To meet this objective, we analyzed several data points in each article published over the last 10 years.The intention was to analyze studies with strong methodological approaches to identify trends of effectiveness.This approach should give providers and families of those suffering from AD a pathway to a non-invasive, cost-effective technique to confirm an AD diagnosis.Findings from this review will hasten practice guidelines for rapid, inexpensive, and accurate diagnosis of AD.

Eligibility Criteria
To be eligible for consideration for this systematic literature review, studies had to use as participants older adults who were undergoing a diagnosis for AD, in a study published in the last ten years, published in a peer-reviewed journal, using strong methodologies.A randomized controlled trial (RCT) was preferred, but we also allowed quasi-experimental, mixed methods, quantitative, and qualitative.A wide variety of interventions were preferred to calculate an overall effect size.This included physical markers, digital markers, and telemedicine interventions.Artificial intelligence interventions such as Bayesian networks and machine learning were also accepted.Machine learning often uses test sets to train the computer in what a positive diagnosis looks like, such as structural equation modeling (SEM).Then, the training is used on a larger set of data to identify the disease.Other systematic reviews were not included in the analysis so as not to confuse the results (because systematic reviews already reported on results from studies that may also be counted in our analysis).Studies must use humans as subjects, be published in the English language, and have a full-text version available for download to enable data extraction.We did not allow gray literature, editorials, protocols, or any other article that did not report study results.This review is registered with PROSPERO: ID 350266.The protocol used for this meta-analysis can be found at Kruse CS.Writing a systematic review for publication in a health-related degree program.JMIR research protocols.2019 Oct 14;8 (10):e15490 [11].

Information Sources
We queried four research databases PubMed (MEDLINE), CINAHL Ultimate (excluding MEDLINE), Web of Science (excluding MEDLINE), and ScienceDirect (excluding MEDLINE).These databases were searched between 25 July 2022 and 26 July 2022.MED-LINE was excluded from all databases except PubMed to help eliminate duplicates.

Search Strategy
We used the Medical Subject Headings (MeSH) of the U.S. Library of Medicine to create a Boolean search string combining key terms listed in the literature: "Alzheimer Disease" AND diagnosis AND (technology OR "artificial intelligence" OR mhealth).We used the same search strategy in all databases and employed the same filters, where available.

Selection Process
We searched for key terms in all databases, filtered results, and screened abstracts for applicability in accordance with the Kruse Protocol [11].At least two reviewers screened all abstracts and rejected any study that did not report results (e.g., protocols, editorials, etc.).Three consensus meetings were held to determine which articles would be analyzed, which data-extraction items were significant measurements of effectiveness, and what observations should become themes.A kappa statistic was calculated.Results were reported in accordance with the PRISMA 2020 standard [12].

Data Collection Process
We used a standardized Excel spreadsheet as a data extraction tool, collecting key data items in each article.This spreadsheet was standardized in the published protocol, and it provides fields of value for administrators, clinicians, and policy makers [11].

Data Items
In accordance with the published protocol, we collected the following fields of data at each step: PICOS (Participant demographics, Intervention specifics, Comparison of results between intervention and control, specific medical Outcomes, and Study design), observed bias, effect size, sensitivity, specificity, F1, country of origin, statistics used, strength of evidence, quality of evidence, measures of effectiveness (cost, efficiency, invasiveness, etc.), and barriers to adoption.A narrative analysis was conducted to identify themes in the literature [13].

Study Risk of Bias Assessment
As each article was analyzed, reviewers noted items of bias (selection bias, sample bias, etc.).We assessed the quality of each study using the John's Hopkins Nursing Evidence Based Practice tool (JHNEBP) [14].Instances of bias helped interpret the results because bias can limit external validity.

Effect Measures
Because reviewers accepted mixed methods and qualitative studies, they were unable to standardize summary measures, as would be performed in a meta-analysis.Measures of effect were summarized in tables where they were reported.Odds ratios, correlation coefficients, and the F1 were converted to Cohen's d [15,16].A weighted average effect size, group specificity, and sensitivity were reported.

Synthesis Methods
Sensitivity, specificity, F1, and effect size were collected from studies.A meta-analysis of this data was performed.Reviewers also performed a thematic analysis of the data collected from studies.Same or similar observations were combined into themes.Themes and other observations were tabulated into affinity matrices.

Reporting Bias Assessment
The overall ratings of quality and strength of evidence, identified using the JHNEBP, provided us with an assessment of the applicability of the cumulative evidence.Reviewers also noted instances of bias such as selection bias, sample bias, and publication bias.

Additional Analyses and Certainty Assessment
A narrative analysis converted observations into themes (common threads between articles) [13].We calculated the frequency of occurrence and reported this in a series of affinity matrices.The frequency reporting provides confidence in the analyzed data.

Study Selection
Figure 1 provides the study selection process to include the inclusion and exclusion criteria from the four research databases.A kappa statistic was calculated based on the levels of agreement between reviewers (k = 0.83, strong agreement).Twenty-eight studies were selected for the systematic literature review.Based on the data provided in these articles, only 15 qualified for the meta-analysis.
collected from studies.Same or similar observations were combined into themes.and other observations were tabulated into affinity matrices.

Reporting Bias Assessment
The overall ratings of quality and strength of evidence, identified using the J provided us with an assessment of the applicability of the cumulative evidence.Re also noted instances of bias such as selection bias, sample bias, and publication b

Additional Analyses and Certainty Assessment
A narrative analysis converted observations into themes (common threads articles) [13].We calculated the frequency of occurrence and reported this in a affinity matrices.The frequency reporting provides confidence in the analyzed d

Study Selection
Figure 1 provides the study selection process to include the inclusion and e criteria from the four research databases.A kappa statistic was calculated base levels of agreement between reviewers (k = 0.83, strong agreement).Twenty-eigh were selected for the systematic literature review.Based on the data provided articles, only 15 qualified for the meta-analysis.

Risk of Bias within and across Studies
We used the JHNEBP quality assessment tool to identify the strength and quality of the evidence.Of the group of articles analyzed, 18/28 (64%) were strength II (quasi-experimental), 6/28 (21%) were strength III (non-experimental, qualitative, or metaanalysis), and 4/28 (14%) were strength I (experimental study, RCT).Additionally, 26/28 (93%) were quality A (consistent results with sufficient sample sizes, adequate control, definitive conclusions), and 2/28 (7%) were quality B (well-defined, reproducible search strategies, consistent results with sufficient numbers of well-defined studies, and definitive conclusions).Most articles analyzed were high-quality with strong methods.
Reviewers also noted instances of bias.The most common bias identified was selection bias.Every study (100%, 28/28) analyzed in this review was conducted in one region of one country, creating a selection bias.Additionally, in 6/28 (21%) of the studies analyzed, instances of sample bias were identified because the sample was heavily skewed with one gender or race.

Results of Individual Studies
Following the Kruse Protocol, reviewers independently recorded observations from each article, commensurate with the objective.Once all reviewers had recorded their observations, a consensus meeting was held to discuss the findings.The result of this meeting was a thematic analysis: making sense of the data.When reviewers noted an observation multiple times, it was recorded as a theme to summarize the observations.These themes are tabulated in Table 2 along with the strength and quality assessments.A match of observation to theme is provided in Appendix A. A summary of other observations can be found in Appendix B.

Results of Syntheses
In 13 articles, effect size was not reported, which eliminated them from the meta-analysis.In 15/28 studies (54%), sufficient data were provided to conduct a meta-analysis.The necessary fields were specificity, sensitivity, F1, odds ratio, Pearson's r, and Cohen's d for effect size.Ten articles reported sensitivity and specificity [17,19,20,26,28,30,33,38,39,42].The overall sensitivity was 0.8516, specificity was 0.8853, precision (PPV) was 0.8388, accuracy was 0.9074, F1 = 0.8452, OR = 44.31CI [35.53, 55.28].The confusion matrix is illustrated in Figure 2. True positives, false negatives, true negatives, and false positives were collected, and odds ratios were calculated.All were converted to Cohen's d [15,16].The weighted average effect size was d = 0.7339, which is a medium effect size that approaches a large effect size.This means the weighted average effect size has a strong medium effect: health information technology interventions can diagnose AD using non-invasive methods.The full table and calculations are listed in Appendix C.

Results of Syntheses
In 13 articles, effect size was not reported, which eliminate analysis.In 15/28 studies (54%), sufficient data were provided to c The necessary fields were specificity, sensitivity, F1, odds ratio, Pe for effect size.Ten articles reported sensitivity and specificity [17,19 The overall sensitivity was 0.8516, specificity was 0.8853, precisio curacy was 0.9074, F1 = 0.8452, OR = 44.31CI [35.53, 55.28].The c trated in Figure 2. True positives, false negatives, true negatives, collected, and odds ratios were calculated.All were converted to weighted average effect size was  ̅ = 0.7339, which is a medi proaches a large effect size.This means the weighted average effe dium effect: health information technology interventions can dia vasive methods.The full table and calculations are listed in Appe

Additional Analysis and Certainty of Evidence
A series of affinity matrices were created to summarize Themes and observations were organized this way to reflect the p

Additional Analysis and Certainty of Evidence
A series of affinity matrices were created to summarize the additional analysis.Themes and observations were organized this way to reflect the probability of their occurrence in the group for analysis.

Interventions of HIT to Diagnose AD
Table 3 summarizes the interventions observed.Four themes and six individual observations were identified by the reviewers for a total of 28 occurrences in the literature.PET was identified in 9/28 (32%) of the studies analyzed [23,27,29,32,36,37,40,41,43].PET biomarkers can identify tau plaque on the brain.This intervention is less invasive, but the confined space of the machine can cause stress.Blood-based biomarkers were identified in 6/28 (21%) of the studies analyzed [17,25,31,34,39,42].This intervention is more invasive than imaging and causes more risk.Various forms of telehealth were implemented to diagnose AD in 4/28 (14%) of the studies analyzed [20,24,33,44].These ranged from video teleconference (VTC) and virtual reality to other televisits to assess cognition.It is important to note that results through telemedicine were the same as traditional visits, which means a clinic can expand productivity without expanding the physical plant, and it exposes the patient to less risk or iatrogenic illness or injury.Multiple forms of artificial intelligence (Bayesian networking, machine learning) were utilized in 3/28 (11%) of the articles analyzed [19,22,28].These interventions utilized PET biomarkers, MRI biomarkers, or arterial spin diagnostics.Six other interventions occurred only once in the literature (21%).These included MRI biomarkers only, MRI and PET biomarkers combined, arterial pulse, RNA, EEG, and spectral domain optical coherence tomography [18,21,26,30,35,38].
Table 5. Barriers commensurate with the adoption of HIT to diagnose AD.
Accurate and early diagnosis of AD is as important to the patient as it is to their family [45] Future research needs to expand some of these interventions into RCTs to firmly establish new CPGs.These new CPGs need to be advertised as widely available to patients and their families.Caring for those with dementia often requires significant changes to lifestyle, habits, and spending.Families need notice to make these adjustments to prepare for their new role and new relationship with the family member.

Limitations
A limitation of this review was the length of time we looked back in the literature (10 years).We chose ten years because it should have identified the most recent advancements in diagnosis intervention.However, a longer time frame may have identified additional interventions.Another limitation is publication bias.We only looked at published works.There are possibly more successful interventions that have not yet been published.Another limitation is that we eliminated editorials and opinion articles because these did not provide empirical evidence.It is possible additional interventions could have been identified from gray literature.

Conclusions
Ten interventions from the literature from the last 10 years were identified as rapid and effective to diagnose AD.While some of these interventions are riskier and involve a blood draw, others are non-invasive and non-stressful to the patient.Additional research needs to expand these interventions into robust RCTs to establish new CPGs for the medical community.
. The patient needs to know about pending secondary effects such as cognitive decline, sleep disruption, or stroke incident to AD.Their families need to know of AD detection to prepare for supportive relationships and quality of life.Providers need to know about the effectiveness of the 10 interventions identified in this research and their relative level of effectiveness (d = 0.7339).The relative effectiveness of these interventions (PET biomarkers, blood-based biomarkers, telehealth, AI/machine learning, MRI biomarkers, PET biomarkers) should lead to new clinical practice guidelines (CPGs).Policy makers need to develop reimbursement mechanisms for the most effective early detection interventions.Many interventions were listed as relatively inexpensive, but medical procedures in general are beyond the out-of-pocket reach of most patients or their care givers.Researchers need to know about early detection interventions to identify more potential participants in the research of this disease to seek treatments and a cure.Many of the interventions are shown to be highly accurate, non-invasive, non-stressful to the patient, repeatable, convenient, and rapid.These considerations can also be utilized to effect change to CPGs in that they show justifiable reasons for the changes.Despite barriers to implementation such as cost, training, physical presence of the patient being necessary, and travel, the aforementioned interventions provide tools to assist in the accurate diagnosis of AD.
MCI versus SCD (AUC, 0.63; p = 0.06).Application of the AD versus SCD discrimination map for prediction of MCI subgroups resulted in good performance for patients with MCI diagnosis converted to AD versus subjects with SCD (AUC, 0.84; p < 0.01) and fair performance for patients with MCI diagnosis converted to AD versus those with stable MCI (AUC, 0.71; p > 0.05)

Table 2 .
Summary of analysis sorted most recent to oldest.

Table 3 .
Interventions of HIT to diagnose AD.

Table 4 .
Effectiveness of intervention of HIT to diagnose AD.