Dementia affects roughly 50 million people worldwide—a figure projected to triple by 2050, according to the World Alzheimer Report. In Australia, accurate counts of people living with dementia remain elusive, hindering effective health-service planning and resource allocation. While routine hospital data capture some dementia diagnoses, many cases go unrecorded, leaving patients at risk of inadequate care. Now, researchers from the National Centre for Healthy Ageing (NCHA)—a collaboration between Monash University and Peninsula Health—have developed a groundbreaking dual-stream algorithm that blends traditional record-coding with artificial intelligence (AI) to dramatically improve dementia detection in hospital electronic health records (EHRs).
The Limitations of Conventional Dementia Coding
Underreporting and Its Consequences
In most Australian hospitals, dementia diagnoses are recorded by medical coders who sift through clinicians’ notes and discharge summaries to apply the appropriate ICD-10 codes. However, the sheer volume of free-text documentation—nursing notes, progress reports, medication orders—makes it difficult to identify every patient exhibiting cognitive decline. Studies suggest that current coding underestimates dementia prevalence by up to 30 percent. This undercount means that individuals may miss out on specialized geriatric assessments, tailored discharge planning and access to community support services. Moreover, policymakers lack a reliable denominator to forecast future care demands.
READ MORE: Personality Profiles May Predict Long‐Term Outcomes in Bipolar Disorder
The Opportunity of Routine Hospital Contact
Hospital admissions present a critical window for dementia detection. Older adults frequently visit emergency departments and inpatient wards, where subtle signs—confusion during triage, medication non-adherence, difficulty providing medical history—may signal underlying cognitive impairment. Yet without systematic approaches, these red flags are often undocumented or overlooked. By harnessing the wealth of structured and unstructured data in EHRs, clinicians and administrators could significantly narrow the dementia-identification gap.
The NCHA’s Dual-Stream AI Approach
Study Design and Cohorts
Lead author Dr Taya Collyer and her team assembled a study cohort of over 1,000 individuals aged 60 and above from the Frankston–Mornington Peninsula region. Half had a specialist-confirmed dementia diagnosis using gold-standard neurocognitive assessments; the remainder served as age-matched controls without cognitive impairment. All participants’ hospital EHRs—spanning admissions, outpatient visits and allied-health encounters—were extracted via the NCHA’s Healthy Ageing Data Platform, an integrated repository of deidentified clinical records.
Stream 1: Structured-Data Algorithm
Building on standard coding methods, the structured-data algorithm incorporated:
• ICD-10 dementia codes and related comorbidity flags (e.g., delirium, Alzheimer’s disease).
• Demographic variables: age, sex, socioeconomic status.
• Medication records: chronic use of anti-dementia drugs (donepezil, memantine) and antipsychotics.
• Utilization patterns: frequent emergency-department visits, repeat admissions for confusion or falls.
• In-hospital events: documented episodes of agitation, wandering or “failure to thrive.”
A machine-learning model (random forest classifier) analyzed these features to assign a dementia-likelihood score for each patient.
Stream 2: Natural Language Processing of Free Text
To tap into the narrative richness of clinical notes, the team deployed a natural language processing (NLP) pipeline:
- Lexicon Development: Clinical experts curated a dictionary of keywords and phrases—“memory lapse,” “disoriented,” “mini-mental status exam,” “cognitive decline”—associated with dementia.
- Text Preprocessing: Notes were deidentified, tokenized and filtered to remove stop-words and irrelevant sections (e.g., administrative boilerplate).
- Contextual Analysis: A transformer-based model (fine-tuned on clinical corpora) parsed sentence structure to differentiate patient-reported symptoms from family history or rule-out assessments.
- Feature Extraction: The algorithm quantified mentions of cognitive symptoms, caregiver concerns, specialist referrals and cognitive-testing results.
The NLP stream yielded its own dementia-probability score, capturing cases that lacked explicit coding but exhibited telltale documentation.
Combining Streams: The Dual-Stream Classifier
The final model integrated outputs from both streams via a logistic-regression meta-classifier, calibrated to maximize sensitivity (true-positive rate) while maintaining high specificity (true-negative rate). Cross-validation on the Frankston–Mornington cohort demonstrated:
• Sensitivity: 92 percent—identifying nearly all patients with confirmed dementia.
• Specificity: 89 percent—minimizing false positives among controls.
• Area Under the ROC Curve: 0.95, indicating excellent overall discrimination.
Validation in a Second Cohort
To ensure robustness, the researchers applied the dual-stream algorithm to a larger external dataset of 2,000 older adults drawn from the U.S. Veterans Health Administration, using a shorter five-factor personality inventory-derived comparator. Despite differences in coding practices and patient demographics, the model maintained sensitivity above 90 percent and specificity around 85 percent—evidence of generalizability across health systems.
Mechanisms of Improved Detection
Early Flags in Structured Data
The structured-data stream quickly flagged patients with repeated delirium episodes or care-home transfers—common harbingers of emerging dementia. However, many older adults present with subtle cognitive symptoms that never reach a diagnostic code until late in the disease course.
Capturing Subtlety in Free-Text Notes
NLP uncovered nuanced descriptions: phrases like “wandered off during lunch” or “difficulty recalling recent conversations” appeared in nursing progress notes, even when coders did not assign a dementia code. In several cases, the model highlighted patients months before formal geriatric referral, suggesting an opportunity for earlier intervention.
Clinical Implications and Future Directions
Enhancing Patient Care in Hospital Settings
With real-time integration into hospital EHR systems, the dual-stream algorithm could flag high-probability cases at the point of care. Geriatric liaison teams could then conduct targeted cognitive assessments, tailor discharge plans—addressing risks such as medication noncompliance and falls—and connect patients with community-based dementia support services.
Informing Health-Service Planning
Accurate counts of persons living with dementia will enable health authorities to project future residential-care needs, workforce requirements and budgetary allocations more precisely. By systematically capturing previously missed cases, the algorithm can serve as an epidemiological surveillance tool for dementia prevalence, informing national estimates and international comparisons.
Ethical and Governance Considerations
Responsible AI Implementation
Lead author Dr Collyer emphasizes the importance of ethical safeguards:
• Data Privacy: All records are deidentified, with strict governance controls on data access.
• Algorithmic Transparency: Clinicians receive explanations of flagging decisions—key phrases and data points that triggered the alert.
• Clinical Oversight: AI outputs supplement, not replace, professional judgment. Final diagnoses remain in the hands of trained specialists.
Training and Trust-Building
Successful deployment requires training clinicians and coders to understand AI-driven flags, interpret probability scores and integrate recommendations into workflows. Engaging multidisciplinary stakeholders—nurses, physicians, allied health professionals—fosters trust and uptake.
Next Steps: Scaling and Continuous Improvement
The research team is collaborating with the Department of Health & Aged Care to pilot the dual-stream algorithm in select Victorian hospitals, monitoring detection rates, patient-outcome metrics and workflow integration challenges. Concurrently, ongoing model refinement will incorporate new data sources—radiology reports, laboratory values and patient-generated health data from wearable devices—to further enhance prediction accuracy.
Conclusion: A Game-Changer for Dementia Detection
As global dementia rates climb, the pressure mounts on health systems to identify and support affected individuals promptly. The NCHA’s novel dual-stream approach—melding structured-data analytics with natural language processing—delivers unprecedented accuracy in detecting dementia from routine hospital records. By harnessing AI responsibly within the Healthy Ageing Data Platform, researchers have laid the groundwork for more timely diagnoses, better‐informed care pathways and robust national surveillance.
Professor Velandai Srikanth, NCHA Director and senior project lead, captures the promise succinctly: “This is a game-changing digital strategy. By capturing clues hidden in free text—descriptions of forgetfulness, confusion or distressed behavior—we can flag patients earlier for the care and support they need. In doing so, we honor our moral and clinical duty to the millions living with dementia in Australia and beyond.”
As AI continues to permeate healthcare, the dual-stream dementia‐detection model stands as a shining example of how technology can amplify human expertise—transforming the identification, counting and management of a condition that touches every community.
