Can artificial intelligence augment COVID-19 screening accuracy?

  • Mei X & al.
  • Nat Med
  • 19 May 2020

  • curated by Liz Scherer
  • Clinical Essentials
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

  • Artificial intelligence (AI) potentially boosts rapidity and screening accuracy for COVID-19 patients who test positive by RT-PCR but have normal CT findings.

Why this matters

  • Augmenting chest CT and clinical information with AI may accelerate COVID-19 screening and diagnosis, especially when testing modalities are unavailable.

Key results

  • 905 patients had chest CT scans.
  • 46.3% (419) tested SARS-CoV-2-positive and 53.7% (486) tested negative by RT-PCR/next-generation sequencing.
    • Negatives confirmed by 2+ RT-PCR tests and clinical observation.
  • Clinical factors: age, exposure history, fever/cough/cough with sputum/white blood cell counts.
  • Joint AI model using clinical data and CT imaging outperformed models using either alone.
  • Joint model vs a senior thoracic radiologist using CT+clinical data (95% CIs):
    • Sensitivity: 84.3% (77.1%-90.0%) AI vs  74.6% (66.4%-81.7%) human.
    • Specificity: 82.8% (75.6%-88.5%) AI vs 93.8% (88.5%-97.1%) human.
    • Area under the curve: 0.92 (0.887-0.948) vs 0.84 (0.800-0.884) human.
  • Test set: 25 patients with COVID-19 and normal chest CT per radiologists at presentation.
    • Deep-learning AI model identified 52% (13/25) scans as COVID-19-positive.
    • Clinical model classified 64% (16/25) as COVID-19-positive.
    • Joint model identified 68% (17/25) as COVID-19-positive.
    • Senior thoracic radiologist, radiologist fellow identified 0% (0/25) as positive.

Study design

  • Evaluation of AI model to augment early detection of SARS-CoV-2 infection.
  • Funding: None disclosed.

Limitations

  • Small numbers.
  • Bias toward patients with COVID-19.