New machine-learning models help spot undetected dementia in primary care

  • Ford E & al.
  • BMC Med Inform Decis Mak
  • 2 Dec 2019

  • International Clinical Digest
Access to the full content of this site is available only to registered healthcare professionals. Access to the full content of this site is available only to registered healthcare professionals.

Takeaway

  • 4 machine-learning models using measures from medical records in the 5 years prediagnosis showed good performance in discriminating between matched primary care patients with and without dementia.

Why this matters

  • Diagnosis of dementia is often missed in primary care.
  • Timely diagnosis allows patients to access specialist assessment, treatment, and support.

 Key results

  • Similarly good discrimination (area under receiver-operating characteristic curve) seen with:
    • Logistic regression model (0.736),
    • Support vector machine model (0.737),
    • Neural network model (0.737), and
    • Random forest model (0.734).
  • Poorest discrimination seen with naive Bayes model (0.682).
  • Top features retained in logistic regression model:
    • Disorientation and wandering,
    • Behavior change,
    • Schizophrenia,
    • Self-neglect,
    • Difficulty managing,
    • Personality change, and
    • Family history of dementia.
  • Highest positive predictive value of 0.312 seen with neural network model.

Study design

  • UK case-control study of 93,120 patients ages >65 (median, 82.6) years.
  • Patients with dementia diagnosis (cases) matched 1:1 by sex, age to patients with no evidence of dementia (controls).
  • 5 machine-learning models developed using 70 clinical measures in 5 years leading up to diagnosis.
  • Main outcome: discrimination between cases and controls.
  • Funding: Wellcome Trust.

Limitations

  • Case-control design, as opposed to cohort design.
  • Need for replication.
  • Reliance on binary clinical measures.