Identifying undetected dementia in UK primary care patients: a retrospective case-control study comparing machine-learning and standard epidemiological approaches

Abstract

BackgroundIdentifying dementia early in time, using real world data, is a public health challenge. As only two-thirds of people with dementia now ultimately receive a formal diagnosis in United Kingdom health systems and many receive it late in the disease process, there is ample room for improvement. The policy of the UK government and National Health Service (NHS) is to increase rates of timely dementia diagnosis. We used data from general practice (GP) patient records to create a machine-learning model to identify patients who have or who are developing dementia, but are currently undetected as having the condition by the GP.MethodsWe used electronic patient records from Clinical Practice Research Datalink (CPRD). Using a case-control design, we selected patients aged >65y with a diagnosis of dementia (cases) and matched them 1:1 by sex and age to patients with no evidence of dementia (controls). We developed a list of 70 clinical entities related to the onset of dementia and recorded in the 5 years before diagnosis. After creating binary features, we trialled machine learning classifiers to discriminate between cases and controls (logistic regression, naïve Bayes, support vector machines, random forest and neural networks). We examined the most important features contributing to discrimination.ResultsThe final analysis included data on 93,120 patients, with a median age of 82.6 years; 64.8% were female. The naïve Bayes model performed least well. The logistic regression, support vector machine, neural network and random forest performed very similarly with an AUROC of 0.74. The top features retained in the logistic regression model were disorientation and wandering, behaviour change, schizophrenia, self-neglect, and difficulty managing.ConclusionsOur model could aid GPs or health service planners with the early detection of dementia. Future work could improve the model by exploring the longitudinal nature of patient data and modelling decline in function over time.

Could dementia be detected from UK primary care patients’ records by simple automated methods earlier than by the treating physician? A retrospective case-control study

“…While recent reviews of dementia risk prediction models show a large body of research on methods to predict likelihood of dementia at a future point 22, 23 , we found few studies that aimed to model automated methods to detect current, early dementia, especially in primary care or using only primary care data 24,25 . So far, there has been little investigation of how accurate an early detection model using only primary care data could be to pick up dementia, or how much earlier than current diagnosis by GPs a tool could discriminate well between patients who go on to be diagnosed and those who do not.…”

“…We constructed a case-control dataset. For this project, CPRD extracted the full records of patients with dementia (cases), identified on the basis of a code list for dementia diagnostic codes (general dementia, vascular and Alzheimer's dementia codes) developed using code lists from Russell et al 27 , and Rait et al 28 , and used in Ford et al 24 , (Appendix 1, see Extended data 29 ). Patients had to be 65 years or older, and had to have records available in CPRD for at least three years prior to the first diagnosis code for dementia, which was recorded between 2000 and 2012.…”

“…We chose a single method to run all analyses on particular subsets of the data by time, and by dementia type (we term these "models"), as our aims were not to assess the effectiveness of different machine learning approaches. Studies suggest random forest algorithms are at least as good as and possibly better than conventional methods such as logistic regression and Cox proportional hazards regression for complex clinical prediction models when binary variables are used 24,33,34 . Our previous research suggested they performed similarly to neural networks and support vector machines on this data source 24 .…”

“…Studies suggest random forest algorithms are at least as good as and possibly better than conventional methods such as logistic regression and Cox proportional hazards regression for complex clinical prediction models when binary variables are used 24,33,34 . Our previous research suggested they performed similarly to neural networks and support vector machines on this data source 24 . They have the additional advantage that the list of features that contribute to the final classification and their relative weights are readily available as a secondary output of the training procedure, giving an indication as to which elements of the patients' EHRs contribute to the classification result most strongly.…”

Can the Use of Bayesian Analysis Methods Correct for Incompleteness in Electronic Health Records Diagnosis Data? Development of a Novel Method Using Simulated and Real-Life Clinical Data

“…be found elsewhere (55) and the full list of variables is given in Appendix 3 (Supplementary Material).…”