The characteristics and performance of health data domains in supporting dementia identification: A systematic review

Abstract

Electronic health records (EHRs) offer growing potential for dementia identification, yet a synthesis of how specific data domains contribute to accurate detection is lacking. This systematic review assessed the role and performance of EHR‐derived data in identifying dementia. Six databases were searched up to April 2025. Fifty studies met inclusion criteria, examining routinely collected health data across care settings. Ten key domains supported dementia identification, including demographics, diagnoses, medications, symptoms, and structured assessments. Most data were sourced from primary care EHRs (48%), with studies primarily conducted in the United Kingdom and United States. Case–control and retrospective cohort designs were commonly applied, often using logistic regression. Cognitive and behavioral domains contributed most to specificity (57.5%–99.9%). Alzheimer's‐specific models had higher accuracy than general dementia models (mean accuracy: 74.6 vs. 67.1). Integrating diverse EHR data, especially cognitive and symptomatic variables, can improve dementia detection. Future research should focus on model validation, standardization, and clinical implementation.

Highlights

• Multi‐domain electronic health record (EHR) data improve dementia identification accuracy.

• Cognitive and behavioral indicators provide highest diagnostic specificity.

• Alzheimer's‐focused models outperform general dementia models.

• Tertiary care EHRs yield more accurate dementia detection.

• Integrating diverse data sources may support earlier dementia diagnosis.

1. BACKGROUND

Dementia represents one of the most pressing public health challenges of the 21st century, characterized by progressive deterioration in cognitive and functional capacity and imposing profound emotional, social, and economic burdens on individuals, families, and healthcare systems worldwide. ¹ Despite its rising prevalence, dementia remains significantly underdiagnosed, ² in part due to the insidious nature of its early manifestations and the limitations of current diagnostic modalities. ³ , ⁴

Information systems have emerged as a transformative force in the way dementia is diagnosed and treated. ⁵ Timely identification of dementia is essential for maximizing the benefits of emerging therapeutics and for initiating supportive interventions that may delay functional decline and mitigate associated burdens. ² , ⁶ However, accurate diagnosis is often complicated by heterogeneous clinical presentations, uncertain etiologies, overlapping symptomatology with normal ageing, and variability in the sensitivity and specificity of diagnostic instruments. ³ , ⁴ Existing tools frequently underperform in detecting early or atypical forms of dementia, particularly among individuals with mild cognitive symptoms or limited healthcare engagement. The failure to recognise dementia as a condition affecting individuals of various age groups and as a condition influenced by modifiable risk factors further complicates the issue. ⁷ Dementia diagnosis is a multifaceted process that involves a comprehensive evaluation of cognitive decline and daily life functioning, corroborated by caregivers or close acquaintances. ⁸ However, this process may be hindered by inconclusive results of existing cognitive screening tools, ⁷ , ⁸ , ⁹ , ¹⁰ where diagnostic sensitivity can vary due to factors such as dementia severity, patient impairment, dementia subtype, and patient–provider contact frequency. ⁸ Additionally, a larger issue can be the failure to identify and assess individuals with milder forms of dementia. This can often occur due to patients and their families being unaware of dementia symptoms or hesitant to seek help ¹¹ , and healthcare providers may miss early signs of dementia, delaying diagnosis.

In response to these challenges, there has been a growing emphasis on using large electronic healthcare databases and predictive modeling to facilitate more effective dementia identification. ¹² These technologies offer a new, potentially powerful platform for enhancing the precision and efficacy of dementia identification and diagnosis. ⁵ , ¹³ , ¹⁴ , ¹⁵ , ¹⁶ , ¹⁷ While the intrinsic value of advanced information systems is the ability to process and analyse extensive patient datasets, ¹⁸ they cannot independently assess individuals or tailor care strategies. The data can, however, be used to identify potential cases of dementia, with human expertise still essential for accurate diagnosis and personalized care planning. Such systems can also enable healthcare professionals to institute timely interventions ¹⁹ and ensure an integrated and more holistic approach to addressing the multifaceted challenges associated with dementia identification and diagnosis (e.g., medication management, appointment scheduling, clinical complexity, diagnostic limitations, systemic and individual barriers, emergence of new subtypes). ²⁰ Here we refer to electronic health record (EHR) systems as an integrated digital platform used by healthcare organizations to capture, store, and manage patient information across clinical encounters. These systems typically include structured data (e.g., diagnostic codes, medication records, laboratory results) as well as unstructured information such as clinical notes or correspondence. Widely used examples include Epic and Oracle Health (formerly Cerner) in the United States, and EMIS Health and SystmOne in the United Kingdom. For the purposes of this review, the term digital health information systems encompass any such digital clinical information EHR system used to support routine care and documentation in primary, secondary, or tertiary settings.

Despite the anticipated benefits of digital health information systems, important knowledge gaps persist. Our understanding of the reliability and applicability of using health information systems to identify dementia across diverse demographic populations remains limited, in particular our understanding of the specific characteristics and performance of health data domains in supporting accurate case identification and diagnoses. To date, only a few reviews have explored the supportive role of digital information systems on dementia identification, and these are largely based in Western countries. ²¹ , ²² , ²³ Variabilities in language and cultural distinctions may influence the portability and fairness of algorithmic models and thus require a more thorough examination to ensure the equitable deployment of technology in dementia diagnosis and care.

There is a need for more studies that systematically examine the reliability, validity, and accuracy of different data elements (e.g., cognitive assessments, neuroimaging, medications) in identifying individuals at risk for dementia, as well as identifying areas for growth for long‐term monitoring and predictive analytics. ⁹ , ¹⁰ , ¹² , ¹³ To address this gap, this systematic review aimed to characterize the data elements used within digital health information systems to support dementia identification, assess their diagnostic performance, and provide evidence‐based guidance for the future development and implementation of digital decision support tools in dementia care.

2. METHODS

This systematic review adheres to the PRISMA guidelines and standards and is registered with PROSPERO (CRD42022356574).

2.1. Inclusion criteria

Studies included in this systematic review met the following criteria: all original research articles and study designs published in peer‐reviewed journals; described the role of digital health information systems in supporting dementia identification across primary, tertiary and community settings using routinely collected healthcare datasets from EHR systems; studies in the English language involving human participants.

2.2. Search strategy

A comprehensive search of 6 electronic databases, including PubMed, Embase, Scopus, PsycINFO, COCHRANE and ProQuest, from inception to September 2023, was conducted by three independent reviewers (J.S., A.B., J.K.), with the search updated in December 2024 and April 2025. The search terms used included, but were not limited to, “dementia,” “Alzheimer's disease,” “health information systems,” “health information technology,” “EHRs,” “identification,” and “diagnosis,” in various combinations, to align with the study's aims and scope (see Supplementary Material). Electronic patient health systems were broadly defined as technology‐based documentation and data collection systems that are used to support the diagnosis and care of individuals with dementia. These systems can have clinical or administrative purposes (e.g., claims or a screening tool), as well as encompassing research systems (e.g., patient registry or an indicator dataset). Other systems (e.g., simulated) were also considered, if relevant. The search strategy was supported by a clinical librarian, an information specialist (J.A.G.) and a machine learning specialist (A.G.). We also hand‐searched reference lists of relevant articles for additional studies.

2.3. Data extraction

Four independent reviewers (J.W., S.G., A.B., J.Z.) screened all abstracts and full‐text articles, resolving any discrepancies through discussion and the assistance of a fourth research member (L.F.L.). Four independent reviewers (J.W., A.B., S.G., J.Z.) extracted data from the selected studies using a standardized data extraction form. Any discrepancies were resolved through discussion and consensus with the lead author (J.S.). The extracted data included study characteristics (e.g., author, year, country, study design, focus, informatic system/platform (e.g., research EHR registries, primary care EHRs, tertiary care EHR), sample type, number of cases, number of controls, mean age, and standard deviation), metrics of systems supporting diagnosis (e.g., clinical factors), and accuracy of diagnosis.

Primary care EHRs were defined as digital platforms designed to manage and organise comprehensive health information for patients within the primary care setting. These systems served as centralized repositories for patient data, including medical history, diagnoses, medications, treatment plans, and communication records. On the other hand, secondary care EHRs commonly used national health insurance database information, including outpatient visits and prescriptions, as well as dementia care and support services. Research EHR registries were specialized databases established by research institutions or collaborative partnerships for the systematic collection, organization, and storage of relevant information for scientific investigations. Data sources for these registries included insurance records and specifically designed databases initiated by research groups. In contrast, tertiary care EHR platforms were designed to manage and organise comprehensive health information for patients receiving advanced medical care. These platforms surpassed the capabilities of primary care EHRs by catering to the complexity inherent in tertiary care. Tertiary care EHRs typically contained more detailed medical records, diagnostic imaging, laboratory results, treatment plans, and other health information relevant to complex medical conditions and interventions.

Domain identification and classification followed a structured, multi‐stage process grounded in qualitative content analysis and EHR phenotyping frameworks. All included studies were initially reviewed in full, and every data element used in model development (e.g., diagnostic codes, cognitive measures, symptoms, functional indicators, laboratory results, medication classes) was extracted verbatim. We captured features exactly as reported by the original study authors, without collapsing or merging at this stage. Three reviewers (J.W., A.B., J.Z.) then independently coded each extracted feature using an inductive approach. Codes reflected the underlying clinical or informational construct (e.g., “memory impairment,” “anticholinesterase prescribing,” “falls,” “neurology referral,” “behavioral disturbance”). No predefined domains were imposed during this stage to avoid bias. After open coding, the reviewers met to group similar codes into broader conceptual clusters. These clusters were grounded in established EHR data categories (diagnostic codes, medication codes, laboratory investigations, clinical assessments) and dementia‐specific symptom domains. This step produced a preliminary set of potential domains (n = 12). Each cluster was defined explicitly with inclusion and exclusion criteria. The preliminary domains were iteratively reviewed by the research team to ensure internal coherence and non‐overlap. Domains capturing conceptually similar information (e.g., general laboratory tests vs pathology) were merged. Conversely, overly broad clusters were separated, for example, behavioral symptoms were separated from cognitive symptoms to reflect distinct clinical phenomena. The three reviewers then independently assigned every extracted feature to one of the refined domains using the operational definitions. Classification decisions were made based on the clinical construct of the feature, its primary purpose in the original study (e.g., diagnostic indicator vs management behavior), and its data type (structured code, narrative data, quantitative measurement). Medication data and diagnostic/procedural codes were grouped within a single domain because many included studies reported these elements together under unified coded clinical data categories. To preserve fidelity to the original reporting and avoid introducing distinctions not supported by the source material, these items were classified within the same domain. Where individual studies provided separable medication and diagnostic information, this is described narratively in the Results section.

Any discrepancies between reviewers were resolved by a fourth reviewer (J.S.). Consensus discussions were guided by the operational definitions and by ensuring that each domain captured a single, conceptually distinct type of information. The final set of ten domains (shown in Supplementary Table S1) thus represent conceptual completeness, minimal overlap between domains, and consistent interpretability across heterogeneous EHR sources. Each domain was then assigned a formal operational definition, scope boundaries, and examples drawn directly from the extracted data. We defined the study size for investigations into positive predictive value (PPV) as the overall count of participants with a dementia code, encompassing both true positives and false positives. For studies examining sensitivity, the study size was determined by the total number of dementia cases in the population, considering both true positives and false negatives as per the reference standard. Where studies reported sufficient numerical detail, we calculated and presented non‐parametric summaries (medians and ranges) to describe variability across domains. However, many included studies provided only counts or proportions without accompanying measures of spread, which limited the feasibility of calculating medians or interquartile ranges for all variables. As a result, summary statistics are presented where possible, and descriptive counts are used when data availability was insufficient.

2.4. Quality assessment

The Mixed Methods Appraisal Tool (MMAT) ²⁴ was used to critically appraise article quality. Two authors (J.K. and A.B.) independently performed the assessments and resolved discrepancies through consensus. Results of the critical appraisal are available in Supplementary Table S1.

2.5. Data synthesis

Due to the heterogeneity of the included studies in terms of study design, information systems, and outcome measures, a narrative, descriptive approach was undertaken. This synthesis aimed to identify patterns, relationships, and themes in the data, and to draw conclusions regarding the role of health data domains in dementia identification.

3. RESULTS

3.1. Overview

A total of 50 eligible articles were identified (Supplementary Figure S1, Table 1). The majority of these studies were conducted in the United Kingdom (21/50, 42%), followed by the United States (14/50, 28%). Study types included case–control (19/50, 38%), retrospective analysis (16/50, 32%), validation (5/50, 10%), randomized controlled trial (2/50, 4%), population‐based cohort (2/50, 4%) and longitudinal cohort (1/50, 2%). Mean sample age ranged from 50 to 85.1 years. Out of the total number of studies, the majority consisted of dementia patients (36/50, 72%), while seven studies focused specifically on non‐dementia patients (7/50, 14%), and a further seven studies investigated both types of patients (7/50, 14%).

TABLE 1.

Summary of study characteristics (n = 50).

					Digital patient record system
Author	Year	Country	Study design	Diagnostic method	Primary care	Secondary care	Tertiary care	Research registry	Sample type	Cases (N)	Controls (N)	Cases mean sample age (SD)	Control mean sample age (SD)
Amra	2017	USA	Validation	Neurological and neuropsychological testing, Dementia Rating Scale		·			PwD	151	151	N/A	N/A
Bamford	2007	UK	Case–control	MMSE, old age psychiatrist notes	·				PwD	319	249	81.3 (6.9)	N/A
Bezabhe	2022	Australia	Retrospective analysis	MedicineInsight dataset, medications	·				All	18,813	N/A	71.9	N/A
Booker	2016	Germany	Case–control	Disease analyser database, ICD‐10 codes	·				PwD	11,956	11,959	80.4 (5.3)	80.4 (5.3)
Brown	2016	UK	Retrospective analysis	ICD‐10 codes	·				PwD	584	866	N/A	N/A
Bucholc	2019	Ireland	Randomized controlled trial	Alzheimer's Disease NeuroImaging Initiative, clinical tests				·	PwD	310	178	N/A	N/A
Burton	2013	UK	Case–control	Primary care consultations, read codes	·				PwD	400	1353	81.4 (6.5)	80.9 (6.2)
Co	2023	UK	Retrospective analysis	ICD‐10		·			PwD	6882	14,493	80.4 (7.0)	82.0 (7.2)
Davies	2011	UK	Case–control	OXMIS/Read codes				·	PwD	9197	39,166	82.2 (7.0)	82.2 (6.8)
Dregan	2015	UK	Case–control	ICD‐10	·				PwD	56,242	350,661	72 (10)	73 (10)
Dunn	2005	UK	Case–control	GP, GP colleague, hospital specialist diagnosis following memory assessment, brain scan or referral	·				PwD	9954	9374	82.1 (6.9)	72.7 (8.0)
Ford	2021	UK	Case–control	Read codes	·				PwD	46,713	46,713	N/A	N/A
Georgiev	2024	UK	Retrospective analysis	Read codes, ICD‐10 codes	·	·			PwoD	11,143	144,113	75	61
He	2024	USA	Validation	ICD‐9/10 codes			·		PwD	2756	N/A	N/A	N/A
Heath	2015	UK	Randomized controlled trial	ICD/10 codes and prescription data			·		PwD	1061	558,942	N/A	N/A
Hsu	2021	Taiwan	Population‐based cohort	ICD‐9/10 codes		·			PwoD	N/A	25,089	N/A	N/A
Huang	2023	Taiwan	Retrospective analysis	ICD‐9/10 codes				·	PwD	206,802	172,335	60 (10)	60 (10)
Imfeld	2012	UK	Case–control	Read codes, prescription	·				PwD	7086	7086	N/A	N/A
Imfeld	2015	UK	Case–control	Read codes, prescription, cognitive tests	·				PwD	16,823	9636	78.8	79.6
Imfeld	2013	Switzerland	N/A	Read codes, cognitive tests				·	PwD	7086	7086	N/A	N/A
Imfeld	2016	UK	Case–control	Read codes	·				PwD	19,463	19,463	N/A	N/A
Jeon	2023	South Korea	Retrospective analysis	ICD‐10 Codes, cognitive tests		·			PwoD	100,282	3,933,382	N/A	55.0 (9.6)
Jeong	2023	Korea	Retrospective analysis	Prescription data, ICD‐10 codes		·			All	11,912	756,469	N/A	52.2 (8.5)
Jick	2000	UK	Case–control	Read codes, prescription				·	PwD	284	1080	N/A	N/A
Kessing	2008	Denmark	Retrospective analysis	ICD‐10 codes				·	PwD	N/A	N/A	N/A	N/A
Khondoker	2023	UK	Retrospective analysis	UK Biobank	·				PwoD	N/A	447,888	N/A	58.0
Kohler	2015	Netherlands	Retrospective analysis	ICPC code P70	·				PwD	978	34,813	N/A	N/A
Lui	2024	USA	Case–control	ICD‐9/10 codes			·		All	23,182	166,931	85.1 (5.3)	N/A
Miled	2020	USA	Validation	ICD‐9/10 codes				·	PwD	2159	11,558	N/A	N/A
Monfared	2023	USA	N/A	TICS‐M, MMSE				·	PwD	11,292	N/A	70.8	N/A
Mukadam	2023	UK	Retrospective analysis	ICD codes, prescription data	·				PwoD	149,228	865,674	N/A	N/A
Nori	2020	USA	Validation	ICD‐9/10 codes, prescription data				·	PwD	N/A	N/A	77	64
O'Malley	2021	UK	Retrospective analysis	ACE‐III, MRI			·		PwD	402	N/A	N/A	N/A
Ramakers	2007	Netherlands	Case–control	DSM‐III	·		·		PwD	74	125	79 (6.2)	79 (7.4)
Ramey	2025	USA	Case–control	SNOMED			·		All	10,552	10.552	N/A	N/A
Regan	2015	UK	Case–control	N/A	·				PwD	153,550	308,843	56 (13)	50 (19)
Rhodius‐Meester	2016	Netherlands	Case–control	MMSE				·	PwD	491	246	N/A	N/A
Salami	2022	Iran	Retrospective analysis	CDR scores				·	PwD	1030	17.527	80.8 (8.2)	80.8 (8.8)
Schliep	2024	USA	Population‐based cohort	ICD codes, census data	·	·			PwoD	522	4208	N/A	76.3 (7.0)
Seshadri	2001	USA	Case–control	Oxford Medical Information System, Prescription Pricing Authority	·				PwD	59	221	66.7	65.2
Shao	2019	USA	Case–control	ICD‐9 codes			·		All	1861	9305	79.8	79.5
Sommerland	2018	UK	Retrospective analysis	ICD‐10 codes, unstructured text			·		PwD	N/A	N/A	82.2	77.9
Soo	2014	Scotland	Validation	ICD‐10 codes	·		·		PwD	N/A	N/A	N/A	N/A
Tang	2024	USA	Case–control	Evaluations at memory clinic			·		All	749	250,545	N/A	N/A
Walker	2018	UK	N/A	Read and product codes, treatment				·	PwD	19,563	19,563	N/A	N/A
Walters	2016	UK	N/A	ICD‐10 codes	·				PwD	284	1080	N/A	N/A
Xu	2020	USA	Retrospective analysis	SNOMED‐CT codes	·				PwD	792	6795	78.4 (5.4)	78.5 (5.4)
Xu	2023	USA	Retrospective analysis	ICD codes	·				All	2501	14,269	74.0 (7.3)	69.9 (9.3)
You	2022	China	Longitudinal cohort	ICD codes, Read codes	·				PwoD	N/A	N/A	N/A	N/A
Yuan	2023	USA	N/A	Cognitive testing, physical/biological assessment, self‐report, MRI, PET		·			PwD	22,553	N/A	N/A	N/A

Abbreviations: All, both cohorts; ICD‐10, International Classification of Diseases, 10^th Revision; ICPC, International Classification of Primary Care; MMSE, Mini‐Mental State Examination; PwD, person with dementia; PwoD, person without dementia.

3.2. Types of EHRs systems

Primary care EHR systems (24/50, 48%) were the most employed data source among researchers. Other databases included research EHR registries (12/50, 24%), tertiary care EHR systems (10/50, 20%), and secondary care EHR systems (8/50, 16%).

3.3. Characteristics of health data domains and methods

Ten different health data domains were identified across different health information systems (Table 2). The domains included were as follows: (1) demographics, such as gender and age (48/50, 96%); (2) medications/medical codes (25/50, 50%); (3) cognitive symptoms (19/50, 38%), including confusion, disorientation, forgetfulness, language difficulties, memory loss, misplacing items, poor judgment, and repetitive actions; (4) physical symptoms (17/50, 34%), examining functional characteristics like medical history, past surgeries, and physical examinations of vision, hearing, and other sensory functions, as well as vital signs; (5) pathology testing and management (15/50, 30%), covering tests, prescribing, and follow‐up sessions; (6) mental health symptoms (13/50, 26%); (7) behavioral symptoms (12/50, 24%), such as aggression, agitation, behavioral changes, personality changes, wandering, and withdrawal; (8) selected tests (7/50, 14%), referring to blood tests, urine tests, and x‐rays; (9) referrals (7/50, 14%); and (10) disturbances of consciousness (6/50, 12%), referring to dizziness, fainting, blackouts, falls, collapses, and sudden lapses. On average, studies included 3.38 health data domains (SD = 2.31, range: 1–10). The most common pattern was using two health data domains.

TABLE 2.

Characteristics of information domains associated with dementia diagnosis

						Type of data extracted by the informatic system
Author	Year	Extraction period	Type of dementia	Case‐detection algorithm	Matching criteria	Demographics	Medications/ medical codes	Cognitive symptoms	Consciousness disturbances	Physical symptoms	Mental health symptoms	Behavioral symptoms	Pathology testing and management	Referrals	Selected tests	Total
Amra	2017	2006–2014	Dementia	Electronic search algorithm	–	·	–	·	–	–	–	–	·	–	–	3
Bamford	2007	1999–2000	Dementia	Logistic regression	Age, gender, general practice	·	·	·	·	·	·	·	·	·	·	10
Bezabhe	2022	2010–2017	Dementia	Cox regression	–	·	–	–	–	–	·	–	·	–	·	4
Booker	2016	2010–2014	Dementia	Logistic regression	Age, gender, health insurance	·	·	·	–	·	·	·	–	–	–	6
Brown	2016	1996–2001	Dementia	N/A	Gender	·	–	–	–	–	–	–	–	–	–	1
Bucholc	2019	N/A	AD	Logistic regression	–	–	–	–	–	–	–	·	–	–	–	1
Burton	2013	2000–2008	Dementia	Logistic regression	Age, gender, general practice	·	–	–	–	–	·	–	·	–	·	4
Co	2023	2008–2017	Dementia	Cox regression	Age, gender, race, ICD codes	·	·	·	–	·	·	–	–	–	–	5
Davies	2011	1997–2008	AD & dementia	Logistic regression	–	·	·	·	–	–	–	–	–	·	–	4
Dregan	2015	1990–2014	Dementia	Logistic regression	–	·	·	–	–	–	–	–	–	–	–	2
Dunn	2005	1992–2002	Dementia	Logistic regression	–	·	–	–	–	–	–	–	–	–	–	1
Ford	2021	2000–2012	Dementia	Logistic regression	Age, gender, general practice	–	·	·	·	·	·	·	–	·	–	7
Georgiev	2024	2009–2023	Dementia	Cox regression	Age, gender	·	·	·	·	·	·	·	–	·	·	9
He	2024	N/A	AD	Computable phenotype	Manual chart review	·	·	·	–	–	–	–	–	–	–	3
Heath	2014	N/A	Dementia	Logistic regression	–	·	–	–	–	–	–	–	·	–	–	2
Hsu	2021	2012–2016	Dementia	Logistic Regression	Age, gender, ICD codes, medications	·	·	·	–	·	·	–	–	–	–	5
Huang	2023	2000–2017	Dementia	Cox regression	–	·	·	·	–	–	–	–	–	–	–	3
Imfeld	2012	1998–2008	Dementia	Logistic regression	–	·	–	–	–	–	–	–	–	·	–	2
Imfeld	2013	1998–2008	AD & dementia	Logistic regression	–	·	·	–	–	–	–	–	–	–	–	2
Imfeld	2015	1998–2013	AD	Logistic regression	–	·	·	–	–	–	–	–	–	–	–	2
Imfeld	2016	1998–2013	AD	Logistic regression	–	·	·	·	·	–	–	–	–	–	–	4
Jeon	2023	2009–2018	AD & dementia	Cox regression	Alcohol use, ICD codes, medication	·	·	–	–	·	–	·	–	–	–	4
Jeong	2023	2009–2018	AD & dementia	Cox regression	Smoking pattern, ICD codes, medication	·	·	–	–	·	–	·	–	–	–	4
Jick	2000	N/A	Dementia	Logistic regression	–	·	–	–	–	–	–	–	·	–	–	2
Kessing	2008	1995–2005	Dementia	N/A	–	·	–	–	–	–	–	–	–	–	–	1
Khondoker	2023	2006–2020	Dementia	Cox regression	–	·	–	–	–	·	·	–	–	–	–	3
Kohler	2015	2000–2012	Dementia	N/A	–	·	–	–	–	–	–	–	·	–	–	2
Liu	2024	2012–2023	AD	Cox regression	Age, gender, race, ICD codes	·	·	–	–	·	–	–	·	–	–	4
Miled	2020	N/A	Dementia	Random forest	Age, gender, race, index date	·	·	–	–	–	–	–	–	–	–	2
Monfared	2023	N/A	AD	Logistic regression	–	·	–	–	–	–	–	–	–	–	–	1
Mukadam	2023	1997–2018	Dementia	Cox regression	Ethnicity	·	–	–	–	·	–	–	–	–	–	2
Nori	2020	2007–2017	Dementia	Recurrent neural network, feed forward network	–	·	·	–	–	–	–	–	–	–	–	2
O'Malley	2022	2018–2021	Dementia	N/A	–	·	·	·	·	·	·	·	·	·	·	10
Ramakers	2007	1996–1999	Dementia	Logistic regression	–	·	–	–	–	–	–	·	–	–	–	2
Ramey	2025	N/A	AD	Multi‐step algorithm	–	·	–	·	–	·	–	–	–	–	–	3
Regan	2015	2002–2013	Dementia	N/A	–	·	–	–	–	–	–	–	·	–	–	2
Rhodius‐Meester	2016	2000–2012	Dementia	PredictAD tool	–	·	–	–	–	–	–	–	·	–	–	2
Salami	2022	N/A	AD	N/A	–	·	–	–	–	–	–	–	–	–	–	1
Schliep	2024	1995–2008	AD	Random forest	–	·	·	–	–	–	·	–	–	–	–	3
Seshadri	2001	1990–1998	AD	Logistic regression	Age	·	–	–	–	–	–	–	·	–	–	2
Shao	2019	N/A	Dementia	Logistic regression	–	·	·	–	–	–	–	–	–	–	–	2
Sommerland	2018	N/A	Dementia	Logistic regression	–	·	–	–	–	·	·	·	–	–	–	4
Soo	2014	N/A	Dementia	N/A	–	·	–	–	–	–	–	–	–	–	–	1
Tang	2024		AD	Random forest	Age, sex, race	·	–	·	–	·	–	–	·	–	–	4
Walker	2018	1997–2015	Dementia	N/A	–	·	–	–	–	–	–	–	–	–	–	1
Walters	2016	2000–2011	Dementia	Cox regression	–	·	·	·	–	–	–	·	–	–	·	5
Xu	2020	2007–2019	Dementia	Logistic regression	–	·	·	·	·	·	·	·	·	·	–	9
Xu	2023	2008–2020	Dementia	Logistic Regression	Drug exposure	·	·	·	–	–	–	–	–	–	–	3
You	2022	2006–2010	Dementia	Logistic regression	–	·	·	·	–	·	–	–	·	–	·	6
Yuan	2023	2005–present	AD	Neural network classification	–	·	–	·	–	–	–	–	–	–	–	2
Total						48	25	19	6	17	13	12	15	7	7

Abbreviations: AD, Alzheimer's disease; ICD, International Classification of Diseases.

Across the 50 included studies, a variety of routinely collected health data sources were used to support dementia identification. The most commonly used data type was diagnostic coding, with ICD‐9 or ICD‐10 codes reported in over half of the studies (26/50, 52%) (Table 1). Other commonly used sources included Read codes, primarily from UK primary care records (13/50, 26%), and prescription or medication data (11/50, 22%), which were often used alongside diagnostic codes to improve case identification. Cognitive and neuropsychological assessments, such as the Mini‐Mental State Examination (MMSE), Addenbrooke's Cognitive Examination (ACE‐III), and the Dementia Rating Scale, were included in 18% of studies (9/50), while clinical documentation from general practitioners, old age psychiatrists, or memory clinics appeared in 12% (6/50). Less frequently used data sources included neuroimaging data such as magnetic resonance imaging (MRI) or positron emission tomography (PET) scans (3/50, 6%), SNOMED‐CT codes (2/50, 4%), and unstructured clinical text (2/50, 4%). Only a small proportion of studies used biological or physical assessments such as lab tests or physical exams (1/50, 2%), self‐reported or survey‐based cognitive data (1/50, 2%), or census and demographic data (1/50, 2%). Additionally, a few studies drew data from specialized research registries, such as the UK Biobank and the Alzheimer's Disease Neuroimaging Initiative (2/50, 4%).

In terms of diagnostic methods, logistic regression case detection algorithms were implemented in almost half of the articles (22/50, 44%), followed by the Cox regression algorithm in 10 instances (10/50, 20%) (Table 2). Other approaches, such as Random Forest (3/50, 6%), bespoke algorithms in electronic searches (1/50, 2%), Recurrent Neural Network, Feed Forward Network (1/50, 2%), Computable Phenotype (1/50, 2%), Multi‐step Algorithm (1/50, 2%), Neural Network Classification (1/50, 2%), and specific instruments such as the PredictAD tool (1/50, 2%) were also employed. Most systems focused on dementia (34/50, 68%), with 12 specifically addressing AD (12/50, 24%) and four examining dementia and AD (4/50, 8%).

3.4. Health data domains and accuracy of dementia identification

Studies with data on accuracy metrics were extracted, resulting in 18 studies. Tables 3, 4, 5, 6 presents accuracy estimates for sensitivity (16/50, 32%), specificity (14/50, 28%), area under the curve (AUC) (8/50, 16%), PPV (7/50, 14%), F‐measure (7/50, 14%), comprehensive overall accuracy (5/50, 10%) and negative predictive value (NPV) (4/50, 8%). Although results varied widely, accuracy levels were modest to high for specificity (57.5%–99.9%), NPV (68.9%–99.8%), AUC (71%–94.4%), and accuracy (73.5%–87.8%) (Table 3). PPV (2.1%–88.0%), F‐measure (3.7%–87.9%), and sensitivity (29.9%–100%) had a larger variation in accuracy across each digital health system (Table 3).

TABLE 3.

Summary of accuracy metrics on articles using information domains to inform dementia diagnosis by digital patient system grouping (tertiary care, secondary care, research registry, combined, primary care)

Digital health system			Accuracy metrics
	Author	Year	Specificity	Sensitivity	PPV	NPV	F‐measure	AUC (95% CI)	Accuracy (Overall correct classification)
TC	He	2024	–	100	37.0–41.5	–	54.1–58.6	–	–
TC	Sommerland	2018	92.0–98.2	50.3–78.0	–	–	–	–	–
SC	Amra	2017	93–99	94–97	–	–	–	–	–
RR	Walker	2018	57.5–79.1	36.0–80.4	–	–	–	–	–
RR	Miled	2020	73.6–80.2	70.9–76.4	–	–	–	–	73.5–79.7
RR	Nori	2020	–	–	–	–	25.6–54.1	80.7–94.4	–
RR	Rhodius‐Meester	2016	66	80	–	–	–	82	–
RR	Bucholc	2019	–	–	–	–	–	–	83.0
RR	Salami	2022	–	81.0–87.8	80.5–88.0	–	79.2–87.9	–	81.0–87.8
PC SC	Georgiev	2025	85.4–89.0	52.8–75.6	14.0–29.2	95.6‐99.1	18‐20	84.7–88.8	–
PC SC	Schliep	2024	74.7–81.9	71.8–80.0	–	–	76.8–77.0	76.9–85.8	–
PC TC	Soo	2014	98.8	29.9	61.5	95.8	–	–	–
PC TC	Ramakers	2007	87	89	–	–	–	–	–
PC	Bamford	2007	68.6–69.8	80.2–81.2	80.9–81.6	68.9–69.0	–	82.3–83.4	76.2–76.8
PC	Ford	2021	91–93	76–85	43–51	–	–	87–94	–
PC	Walters	2016	73.1–99.9	72.0–77.7	2.1–10.9	99.3–99.8	–	–	–
PC	Xu	2020	74.7–86.8	60.2–72.7			67.6–74.4	71.0–76.4
PC	You	2022	75.6–76.4	77.4–81.5	–	–	3.7–7.5	84.7–89.0	75.6–76.4

Abbreviations: AUC, area under the curve; NPV, negative predictive value; PC, primary care; PPV, positive predictive value; RR, research registry; SC, secondary care; TC, tertiary care.

TABLE 4.

Specificity outcomes by study and inclusion of health data domains

Study	Type of dementia	Specificity range	Demographics	Medications/Medical codes	Cognitive symptoms	Consciousness disturbances	Physical symptoms	Mental health symptoms	Behavioral symptoms	Pathology testing and management	Referrals	Selected sample tests
Walker 2018	Dementia	<60	Yes	No	No	No	No	No	No	No	No	No
Rhodius‐Meester 2016	Dementia	60–70	Yes	No	No	No	No	No	No	Yes	No	No
Bamford 2007	Dementia	60–70	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Miled 2020	Dementia	70–78	Yes	Yes	No	No	No	No	No	No	No	No
Schliep 2024	AD	70–78	Yes	Yes	No	No	No	Yes	No	No	No	No
Walters 2016	Dementia	70–78	Yes	Yes	Yes	No	No	No	Yes	No	No	Yes
Xu 2020	Dementia	70–78	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No
You 2022	Dementia	70–78	Yes	Yes	Yes	No	Yes	No	No	Yes	No	Yes
Ramakers 2007	Dementia	80–90	Yes	No	No	No	No	No	Yes	No	No	No
Georgiev 2025	Dementia	80–90	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes	Yes
Sommerland 2018	Dementia	90+	Yes	No	No	No	Yes	Yes	Yes	No	No	No
Ford 2021	Dementia	90+	No	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes	No
Amra 2017	Dementia	90+	Yes	No	Yes	No	No	No	No	Yes	No	No
Soo 2014	Dementia	90+	Yes	No	No	No	No	No	No	No	No	No

TABLE 5.

Sensitivity outcomes by study and inclusion of health data domains

Study	Type of dementia	Sensitivity range	Demographics	Medications/ Medical codes	Cognitive symptoms	Consciousness disturbances	Physical symptoms	Mental health symptoms	Behavioral symptoms	Pathology testing and management	Referrals	Selected sample tests
Soo 2014	Dementia	<60	Yes	No	No	No	No	No	No	No	No	No
Walker 2018	Dementia	<60	Yes	No	No	No	No	No	No	No	No	No
Sommerland 2018	Dementia	<60	Yes	No	No	No	Yes	Yes	Yes	No	No	No
Xu 2020	Dementia	60–70	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No
Georgiev 2025	Dementia	70–78	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes	Yes
Miled 2020	Dementia	70–78	Yes	Yes	No	No	No	No	No	No	No	No
Walters 2016	Dementia	70–78	Yes	Yes	Yes	No	No	No	Yes	No	No	Yes
You 2022	Dementia	70–78	Yes	Yes	Yes	No	Yes	No	No	Yes	No	Yes
Ford 2021	Dementia	70–78	No	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes	No
Schliep 2024	AD	70–78	Yes	Yes	No	No	No	Yes	No	No	No	No
Salami 2022	AD	80–90	Yes	No	No	No	No	No	No	No	No	No
Rhodius‐Meester 2016	Dementia	80–90	Yes	No	No	No	No	No	No	Yes	No	No
Ramakers 2007	Dementia	80–90	Yes	No	No	No	No	No	Yes	No	No	No
Bamford 2007	Dementia	80–90	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Amra	Dementia	90+	Yes	No	Yes	No	No	No	No	Yes	No	No
He 2024	AD	90+	Yes	Yes	Yes	No	No	No	No	No	No	No

TABLE 6.

Accuracy outcomes (PPV, NPV, F‐measure, AUC and overall classification) by study and inclusion of health data domains

Study	Type of dementia	Metric	Range	Demographics	Medications/ Medical codes	Cognitive symptoms	Consciousness disturbances	Physical symptoms	Mental health symptoms	Behavioral symptoms	Pathology testing and management	Referrals	Selected sample tests
Walters 2016	Dementia	PPV	<60	Yes	Yes	Yes	No	No	No	Yes	No	No	Yes
Georgiev 2025	Dementia	PPV	<60	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes	Yes
Ford 2021	Dementia	PPV	<60	No	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes	No
He 2024	AD	PPV	<60	Yes	Yes	Yes	No	No	No	No	No	No	No
Soo 2014	Dementia	PPV	60–70	Yes	No	No	No	No	No	No	No	No	No
Bamford 2007	Dementia	PPV	80–90	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Salami 2002	AD	PPV	80–90	Yes	No	No	No	No	No	No	No	No	No
Bamford 2007	Dementia	NPV	60–70	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Georgiev 2025	Dementia	NPV	90+	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes	Yes
Walters 2016	Dementia	NPV	90+	Yes	Yes	Yes	No	No	No	Yes	No	No	Yes
Soo 2014	Dementia	NPV	90+	Yes	No	No	No	No	No	No	No	No	No
You 2022	Dementia	F‐measure	<60	Yes	Yes	Yes	No	Yes	No	No	Yes	No	Yes
He 2024	AD	F‐Measure	<60	Yes	Yes	Yes	No	No	No	No	No	No	No
Georgiev 2025	Dementia	F‐Measure	<60	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes	Yes
Nori 2020	Dementia	F‐measure	<60	Yes	Yes	No	No	No	No	No	No	No	No
Xu 2020	Dementia	F‐measure	60–70	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No
Salami 2022	AD	F‐measure	70–78	Yes	No	No	No	No	No	No	No	No	No
Schliep 2024	AD	F‐measure	70–78	Yes	Yes	No	No	No	Yes	No	No	No	No
Xu 2020	Dementia	AUC	70–78	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No
Bamford 2007	Dementia	AUC	80–90	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
Ford 2021	Dementia	AUC	80–90	No	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes	No
Georgiev 2025	Dementia	AUC	80–90	Yes	Yes	Yes	Yes	Yes	Yes	Yes	No	Yes	Yes
Rhodius‐Meester 2016	Dementia	AUC	80–90	Yes	No	No	No	No	No	No	Yes	No	No
Schliep 2024	AD	AUC	80–90	Yes	Yes	No	No	No	Yes	No	No	No	No
You 2022	Dementia	AUC	80–90	Yes	Yes	Yes	No	Yes	No	No	Yes	No	Yes
Nori 2020	Dementia	AUC	80‐90	Yes	Yes	No	No	No	No	No	No	No	No
Bamford 2007	Dementia	Overall	70–78	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes	Yes
You 2002	Dementia	Overall	70–78	Yes	Yes	Yes	No	Yes	No	No	Yes	No	Yes
Miled 2020	Dementia	Overall	70–78	Yes	Yes	No	No	No	No	No	No	No	No
Salami 2022	AD	Overall	80–90	Yes	No	No	No	No	No	No	No	No	No
Bucholc 2019	AD	Overall	80–90	No	No	No	No	No	No	Yes	No	No	No

Abbreviations: AD, Alzheimer's disease; AUC, area under the curve; NPV, negative predictive value; PPV, positive predictive value.

The 18 studies that had accuracy metrics reported (i.e., on specificity, sensitivity, PPV, NPV, F‐Measure, AUC, overall correct classification accuracy) mainly focused on using demographics (16/18, 89%) information as a key diagnostic factor, followed by medications/medical codes (10/18, 56%), behavioral symptoms (8/18, 44%), cognitive symptoms (8/18, 44%), physical symptoms (6/18, 33%), mental health symptoms (6/18, 33%), pathology testing and management (5/18, 28%), disturbances in consciousness (4/18, 22%), referrals (4/18, 22%), and selected tests (4/18, 22%).

Various factors impacted the specificity metrics for dementia diagnosis. We recategorized results based on accuracy ranges (i.e., < 60%, 60%–70%, 70%–78%, 80%–90%, and +90%) (Table 4). The inclusion of medications/medical codes, cognitive symptoms and pathology testing and management health data domains tended to generate medium specificity (60%–78%).

In the context of the higher accuracy ranges (80%–90% and +90%), the presence of cognitive, physical, mental health, and behavioral symptoms most commonly supported the accuracy of diagnostic specificity (Table 4). There was no clear relationship between multiple data domain inclusion and higher sensitivity; however, studies with medium sensitivity outcomes (60%–78%) generally included medications/medical codes, cognitive, physical, and behavioral symptoms (Table 5). For PPV and NPV measures, there was a mix of variables across all performance outcomes (i.e., medications/medical codes were found in both low and high outcomes) (Table 6). Studies that included more than five health domains tended to have lower PPV values (< 60). F‐measures ranged from < 60% to 78% across all seven reporting studies (Table 6). There was no distinct pattern between health domains and higher performance on the F‐measure. Most studies had high AUC (80%–90%, 7/8) (Table 6). There was no distinct health data domain that was associated with high or low AUC. The inclusion of a single variable was associated with higher (80%–90%) overall correct classification (i.e., Salami 2022, Bucholc 2019) (Table 6). The inclusion of two or more health data domain factors was linked to slightly lower (70%–78%) overall correct classification (e.g., Bamford 2007, 10 factors; You 2002, 6 factors).

The majority of the studies with accuracy measures focused specifically on dementia (14/18, 78%). Four studies focused on AD (4/18, 22%), and no studies included both dementia and AD. Only one AD study (Schliep, 2024) reported on specificity. In sensitivity outcomes, AD‐focused studies performed better (2/3 studies reported 80%+) than dementia‐focused studies (9/13 studies reported <60%–78%) (Table 6). Only dementia‐focused studies reported on NPV. No discernible differences were observed between AD and dementia within PPV and AUC metrics. AD‐focused studies had higher F‐measure metrics on average (two of three studies reported 70%–78%) as compared to dementia‐focused studies (three of four reported < 60%) (Table 6). This pattern was similar for overall correct classification (AD‐focused studies reported 80%–90% vs dementia‐focus studies 70%–78%) (Table 6).

4. DISCUSSION

The capacity of digital health systems to support accurate and timely diagnosis of complex conditions such as dementia is increasingly regarded as a cornerstone of contemporary health system transformation. Our review identifies high levels of diagnostic performance across multiple algorithmic approaches, with many studies reporting sensitivity, specificity, and overall classification accuracy exceeding 80%. These findings reaffirm the promise of data‐driven methods in augmenting diagnostic decision‐making, yet they also raise essential questions about the architecture, integrity, and contextual appropriateness of such systems when embedded within clinical practice.

A salient observation from this review is the centrality of demographic information (age, sex, and socioeconomic status) in predictive models. The near‐universal inclusion of demographic variables (93.9%) is unsurprising given their routine availability in EHRs and their well‐documented associations with dementia risk. ²⁵ , ²⁶ , ²⁷ These variables, while conceptually straightforward, can act as powerful proxies for social determinants of health. ²⁸ Socioeconomic disadvantage, for instance, has been repeatedly linked to disparities in cognitive health trajectories through mechanisms involving education, comorbidities, access to care, and cumulative stress exposure. ²⁹ , ³⁰ , ³¹ It is imperative, however, that such variables not be treated as static artifacts. Rather, they should be interpreted within the evolving life‐course contexts of individuals, ³² informing not only risk stratification but also equity‐oriented design in diagnostic systems.

The existing literature emphasises that diverse demographic profiles may predispose individuals to varying health risks (e.g., higher rates of chronic diseases, lower life expectancy, depression, cognitive decline ²⁷ ). Socioeconomic factors, in particular, can impact access to healthcare resources and influence lifestyle choices which contributes to variations in health outcomes. These findings indicate perhaps the basic need for health systems to integrate demographic considerations into diagnostic protocols to better support our contextual understanding of individuals and provide appropriate support.

Medication history was also a potential predictor of diagnostic accuracy; however, this factor may reflect the bidirectional link between pharmacological burden and cognitive function. Research has extensively documented medication‐related cognitive impairment, with certain drug classes associated with increased cognitive risk. ³³ , ³⁴ , ³⁵ Drugs with anticholinergic properties, in particular, are consistently associated with cognitive decline, ³⁶ and are both readily extractable from EHRs and modifiable in clinical practice. The integration of medication data into diagnostic algorithms thus represents a low‐barrier, high‐utility enhancement to predictive modeling. From a systems design perspective, these findings argue for the standardization and semantic interoperability of medication data fields across health information systems, which can enable machine‐readable interpretations that can dynamically inform risk estimation. ³⁷ , ³⁸

The diagnostic significance of cognitive and behavioral symptom domains further emphasizes the need for multimodal and temporally sensitive data capture. Traditional neuropsychological assessments, while foundational, often lack ecological validity and are subject to under‐recording in real‐world EHRs. ³⁹ Our findings suggest that digital systems capable of integrating structured behavioral observations, cognitive test results, and longitudinal symptom trajectories are better positioned to support early detection. This approach aligns with contemporary shifts toward multi‐dimensional diagnostic frameworks and calls for the operationalization of structured templates within EHRs to facilitate consistent capture of cognitive‐behavioral data in primary and secondary care.

From a methodological standpoint, one of the most intriguing findings relates to the performance of reduced‐variable models. Several studies explored the performance of reduced‐variable models to indicate that streamlined feature sets can, in some contexts, achieve accuracy comparable to, though not necessarily exceeding, that of more complex models. ²² , ⁴⁰ , ⁴¹ , ⁴² , ⁴³ , ⁴⁴ , ⁴⁵ For instance, Haro Alonso et al. ⁴⁴ reported that a six‐feature machine learning model performed similarly to a logistic regression model using fourteen features, while also illustrating the trade‐offs involved, including potential reductions in prognostic precision. Other studies examining reduced or minimum variable sets noted that, for certain use‐cases, limiting the number of predictors did not substantially degrade model performance and occasionally improved stability by reducing noise. ⁴³ , ⁴⁶ , ⁴⁷ Evidence on the contribution of imaging variables was mixed; while some studies found that removing specific imaging features modestly improved accuracy, others reported benefits from combining imaging and clinical information. These disparities again reflect the heterogeneity in data quality, modeling strategy, and outcome definitions. ⁴⁵ However, recent research has indicated that simplified machine learning models using a reduced set of features demonstrate higher prognostic accuracy compared to traditional logistic regression and Cox proportional hazards models that incorporate clinical and imaging variables. ⁴³ , ⁴⁶ , ⁴⁷ These findings should be interpreted cautiously given the diversity of study designs and evaluation metrics. The collective evidence suggests that dimensionality reduction and careful feature selection may support model interpretability and feasibility, especially in settings with incomplete data, rather than demonstrating intrinsic superiority of reduced models or machine learning approaches over traditional statistical models. ⁴⁸ , ⁴⁹ The capacity of some machine learning methods to generate patient‐level explanations represents a potential advantage, but this remains contingent on model type, data quality, and validation approach.

Across included studies, performance differences between machine learning approaches and logistic regression were inconsistent. Variability in feature engineering, sample size, and outcome definition appears to exert greater influence on performance than model class itself. Our review therefore does not support strong conclusions about inherent superiority of one modeling approach. Hybrid approaches, where interpretable features are embedded within more complex architectures, may offer a pragmatic path forward, to balance technical performance with clinical usability. ⁵⁰

4.1. Implications

Based on observed patterns, several domain combinations appear promising for different phenotyping objectives. Algorithms incorporating diagnostic codes together with medication records (particularly cholinesterase inhibitors), cognitive assessment results, and neuroimaging reports tended to yield higher specificity in studies where these data were available. Combining structured (codes) and semi‐structured domains (narratives, cognitive scores) may help differentiate AD from other dementias. On the other hand, broader combinations including diagnostic codes, healthcare utilization indicators, and functional/behavioral assessments may enhance sensitivity, particularly when the goal is case ascertainment within large health systems. Yet, for high‐specificity use‐cases (e.g., cohort enrichment), narrower rule‐based phenotypes emphasizing diagnostic codes plus confirmatory domains (medications, cognitive assessments) may be appropriate. For surveillance or screening, wider inclusion of utilization patterns, laboratory values, or narrative features may help capture early or atypical presentations. However, these recommendations remain provisional given the descriptive nature of the evidence and should be refined through future head‐to‐head evaluations.

Regardless, at a systems level, our review reinforces the imperative to reframe EHRs not just as repositories of administrative data but as diagnostic platforms capable of real‐time inference. The routine and structured capture of demographic, pharmacological, cognitive, and behavioral data should be considered a minimum design standard in contemporary health information systems. Furthermore, the integration of diagnostic algorithms into clinical workflows will require interoperability across care settings, explainable outputs, and interfaces that facilitate clinical reasoning rather than obscure it.

Digital diagnostics must also grapple with the tension between comprehensiveness and feasibility. Although large‐scale models incorporating a wide array of variables may appear to offer greater fidelity, they often falter in real‐world environments where data completeness is variable and time for decision‐making is constrained. Our findings support a “minimum viable model” paradigm, where diagnostic tools are built around a core set of high‐value, clinically actionable variables. This approach not only aligns with the principles of lean health informatics but also mitigates the risk of overfitting and enhances generalizability.

4.2. Limitations

Our review is limited by the heterogeneity of included studies, which precluded quantitative synthesis and limits comparative inference. Definitions of the diagnostic endpoint varied considerably; while some studies used formal diagnostic criteria (e.g., DSM), others relied on clinical codes or cognitive test thresholds, and this thus raises concerns about the comparability and reliability of the reference standard. The predominance of cross‐sectional designs also limits insight into the prognostic utility of these models over time, as well as the large majority of studies being in Western countries with limited diversity. Furthermore, the absence of rigorous reporting on data quality, especially within EHR‐based studies, raises the possibility that reported accuracies may be overestimated in high‐fidelity research environments and underperform in practice.

5. CONCLUSION

This review highlights the evolving capabilities of digital health systems to support accurate dementia identification through the intelligent integration of routinely collected health data. While demographic, medication, and symptom‐based data domains contribute substantively to predictive performance, their value is contingent upon data quality, structured capture, and context‐appropriate use. Reduced and optimized diagnostic models demonstrate substantial promise, particularly when aligned with health system capabilities and clinical workflows. Future work should focus on longitudinal validation, implementation science, and the co‐design of decision‐support tools that enhance, not disrupt, clinician cognition. The path forward requires not only better models but also better systems.

AUTHOR CONTRIBUTIONS

Ali Butt: Investigation; writing—review and editing. Andrew Georgiou: Conceptualization; writing–review and editing. Anupama Ginige: formal analysis; visualization; writing—review and editing. Jiwoo Kim: investigation; formal analysis; writing—original draft; writing—review and editing. Lee‐Fay Low: Conceptualization; writing—review and editing. Joyce Siette: Conceptualization; methodology; investigation; formal analysis; writing—original draft; writing—review and editing; supervision. Sandra Sonego: Writing—original draft; writing—review and editing. Jessica‐Lee Zammit: Investigation; formal analysis; writing— review and editing.

CONFLICT OF INTEREST STATEMENT

The authors declare no conflicts of interest. Author disclosures are available in the Supporting Information.

Supporting information

Supporting Information

Supporting Information

Supporting Information