Artificial Intelligence in Medical Diagnosis

Abstract

Background *
Rapid advances in machine learning and deep learning have expanded the diagnostic capabilities of artificial intelligence (Al) across radiology, pathology, and primary care. Despite promising accuracy gains and workflow efficiencies, questions remain about generalizability, bias, and clinician acceptance.

Methods *
Rapid advances in machine learning and deep learning have expanded the diagnostic capabilities of artificial intelligence (Al) across radiology, pathology, and primary care. Despite promising accuracy gains and workflow efficiencies, questions remain about generalizability, bias, and clinician acceptance.

Results *
42 primary studies (11 pathology, 8 primary care, and 23 radiology) satisfied the inclusion criteria. The median number of cases in the test set was 6,800 (IQR 2,400-15,200). In radiology studies, Al's pooled AUC was 0.92 (95% CI 0.88-0.95), while clinicians' AUC was 0.89 (95% CI 0.85- 0.93). With no discernible loss of specificity, Al- assisted slide review increased the sensitivity of mitosis detection in pathology by 14%. On average, primary care decision support systems shortened diagnostic turnaround times by 22%. Nevertheless, only 29% of studies addressed algorithmic bias or explainability, and 64% of studies lacked external validation.

Conclusion *
Al is poised to play a transformative role in medical diagnosis by enhancing accuracy, speeding up processes, and improving patient outcomes. However, ethical deployment, careful validation, and collaborative approach between Al systems and healthcare professionals are essential for its responsible integration into medical practice. The future of Al in diagnosis is promising but must be navigated with caution and compassion.