Development and validation of clinical prediction models for dementia identification in non-specialist settings

A timely dementia diagnosis is a public health priority. However, there is no single accurate test suitable for the identification of dementia in non-specialist settings, such as primary care. The aim of this thesis was to develop and validate clinical prediction models for a computerised clinical decision support system being developed (DECODE). The DECODE system estimates the probability of dementia for a given patient, and provides clinical recommendations such as whether a full dementia evaluation is appropriate. The thesis includes a narrative review, an investigation of dementia misclassification, and a series of model development and validation analyses. The narrative review examined case-finding policy and practice in the UK and US. An evidence-based pathway for dementia case-finding was proposed, informing the rationale for and design of DECODE. In the investigation of dementia misclassification three brief cognitive assessments were analysed in a population-based cohort of older adults. Misclassification ranged from 14-21% across the assessments, each associated with different patient characteristics. Only 3% of participants were misclassified by all three assessments, suggesting misclassification may occur due to test-specific biases. Prediction model development began with identification of candidate predictors of dementia based on dementia-related systematic reviews, diagnostic criteria and expert judgement. A total of 40 candidate predictors included socio-demographics, subjective and objective cognition, functional impairment, family history and health factors available in non-specialist settings. A series of bootstrapped fractional polynomial logistic regression analyses predicted current dementia status using two population-based cohorts from the US (Aging, Demographics, and Memory Study, N = 856) and Australia (Sydney Memory and Ageing Study, N = 707). Models were developed and internally evaluated for use with four different brief cognitive assessments; the Mini-Mental State Examination (MMSE), Memory Impairment Screen, General Practitioner Assessment of Cognition and 10-point Cognitive Screener. Final models were externally validated using the US National Alzheimer’s Coordinating Center (NACC) memory clinic dataset (N = 27,235). There were no differences in area under the curve (AUC) between internal and external validation. All models were consistently more accurate than brief cognitive assessments alone (difference in AUC all p<.001, AUCs ranging from 0.93 to 0.97 depending on incorporation of cognitive assessment results). In the external validation, the overall performance of the DECODE model incorporating the MMSE (AUC = 0.98, 95% CI = 0.97 – 0.98) was greater than that for the MMSE alone (AUC = 0.93, 95% CI = 0.91 – 0.95), p<.001. The MMSE alone misclassified 13.6% of patients, whereas the DECODE model misclassified 10.1%, reducing misclassification by 25.7%. As a further extension of the analyses an alternative version of DECODE was developed incorporating a longer cognitive assessment, the Montreal Cognitive Assessment (MoCA), with potential application in memory clinics. Performance of the MoCA version of DECODE was superior to those incorporating brief cognitive assessments (AUC = 0.99). In conclusion, DECODE clinical prediction models are able to accurately identify dementia status in a variety of clinical contexts and outperform conventional brief cognitive assessments. Further validation and clinical utility testing is required to assess whether these algorithms incorporated into the DECODE computerised clinical decision support system have the potential to improve the diagnostic pathway for dementia.The Halpin Trus