Objective
To evaluate whether PurpleAI’s HyperInsight–ICH v2.0.1 could improve the prioritization of acute intracranial hemorrhage cases without reducing radiologists’ diagnostic performance.
Study Design
- Retrospective, single-center, pre/post implementation study
- 956 non-contrast brain CT examinations
- Pre-AI period: 541 scans over 35 days
- Post-AI period: 415 scans over 35 days
- AIH prevalence: approximately 13%
- Two board-certified general radiologists interpreted the scans in routine clinical practice.
- Two experienced neuroradiologists independently established the reference standard.
Key Results
| Outcome | Before AI | After AI | Significance | | ----------------------------------------------------- | --------- | -------- | --------------------- | | Median reading position of AIH cases | 7.25 | 1.5 | P < 0.001 | | Chronological vs. actual reading order, Kendall’s Tau | 0.61 | 0.01 | P < 0.001 | | Radiologist accuracy | 0.99 | 0.99 | No significant change | | Radiologist sensitivity | 1.00 | 1.00 | No significant change | | Radiologist specificity | 1.00 | 0.99 | No significant change |
The proportion of hemorrhage cases reviewed within the first quarter of the worklist also increased significantly. HyperInsight–ICH achieved standalone AUC 0.99, sensitivity 0.96, and specificity 0.99 during the post-implementation period.
Interpretation
The study’s strongest finding is not improved diagnostic accuracy, which was already very high. It is that AI-based worklist integration helped radiologists review suspected hemorrhage cases substantially earlier without compromising performance.
Raw turnaround time also decreased, but it improved for both hemorrhage and non-hemorrhage cases. Therefore, the study cannot attribute the entire turnaround-time reduction to AI.
Limitations
- Retrospective, non-randomized pre/post design
- Single institution and one CT manufacturer
- Only two radiologists
- Short 70-day study period
- No demonstrated improvement in treatment time or patient outcomes
- Prospective, multicenter validation remains necessary
Editorial note: The abstract and main results table report slightly different early-diagnosis-rate values. The statistically significant direction of improvement is consistent, but the exact figure should be used cautiously.
Journal: Diagnostic and Interventional Radiology
Published: July 7, 2025 | DOI: 10.4274/dir.2025.253301