ICH AI Triage: Time-to-Read Reduction and the Evidence Behind It

Intracranial hemorrhage triage is the indication with the deepest published evidence base for radiology AI clinical benefit. Aidoc's ICH detection system (K183327) has been evaluated in multiple published studies examining time-to-read outcomes, and findings consistently show that AI-assisted flagging reduces the interval between study completion and radiologist first read for ICH-positive cases. The mechanism is straightforward: the AI system identifies the study as high-priority and either moves it up the worklist or generates an alert, causing the radiologist to read it before reaching it in chronological order.

But "time-to-read reduction" as a headline metric requires careful examination before a community hospital makes a procurement decision based on it. What exactly is being measured, what is the magnitude of the reduction, and does the study population match the community hospital environment? This article examines those questions directly.

What "Time-to-Read" Actually Measures

Time-to-read in published AI triage studies is typically defined as the interval from CT acquisition (or PACS arrival) to radiologist dictation or first access of the study in the PACS viewer. It is not time-to-treatment, time-to-evacuation, or time-to-discharge. The distinction matters because the downstream clinical impact of reading a study earlier depends on what happens after the read — neurosurgery availability, ICU capacity, transfer protocols — factors that are not controlled by the radiology AI system.

Published ICH triage studies, including analyses from NEJM Evidence and JACR-published retrospective reviews, show time-to-read reductions for ICH-positive studies in the range of 30–96 minutes depending on baseline workflow conditions, shift configuration, and deployment environment. The larger reductions tend to occur in overnight shifts with low radiologist coverage — exactly the community hospital scenario — because there is more slack in the baseline queue that the AI can exploit by moving the critical study to the front.

The Community Hospital ICH Context

Consider the typical community hospital overnight scenario in detail. A 150-bed regional facility in rural New England staffs one teleradiology reader overnight via contract. Studies queue in PACS via DICOM C-STORE from the CT scanner in the ED and from inpatient CT. The teleradiologist is reading remotely, working through a RIS worklist sorted by arrival time, covering this facility alongside two or three others simultaneously.

When a CT head study arrives at 2:17 AM with an intraparenchymal hemorrhage — a hypertensive bleed in the basal ganglia, 8 ml by ABC/2 method — it enters the queue in position 12, behind a mix of chest CTs, a follow-up knee MRI, and routine head CT for altered mental status without acute finding. Without triage, the ICH study will be read when the teleradiologist reaches it in sequence, potentially 45–90 minutes after acquisition if the queue is long. With AI triage and worklist re-ranking, that study moves to position 1 or 2 based on its triage score, and the teleradiologist's next action is to open the ICH-positive study.

The time-to-read reduction in this scenario is real and clinically consequential. A neurosurgical consultation for expanding ICH, a transfer decision to a facility with neurosurgical capability, or simply a call to the ED team that changes the monitoring level — all of these downstream actions are gated on the radiologist's read. Moving that read 60 minutes earlier is meaningful.

Methodological Limitations in Published ICH Triage Evidence

We're not saying the published ICH triage evidence overstates the benefit. We're saying there are methodological considerations that affect how directly these numbers translate to your specific environment.

Most published studies are retrospective observational designs comparing time-to-read before and after AI deployment at a single site or small number of sites. They are not randomized controlled trials — which would be difficult to design ethically for this indication since withholding a triage tool from the control arm involves deliberate delay of care for a life-threatening finding. The observational design means baseline workflow drift, seasonal staffing variation, and concurrent quality improvement initiatives can confound the before/after comparison.

Additionally, published studies are disproportionately from academic or large community facilities with dedicated radiology informatics teams who manage the AI deployment carefully, monitor performance, and recalibrate thresholds when needed. Smaller community hospitals without this infrastructure may see different results — potentially better if their baseline workflow is more chaotic and the ceiling for improvement is higher, or worse if the deployment isn't maintained properly.

ICH subtype also matters for performance. Aidoc K183327 and similar cleared systems have strongest published evidence for intraparenchymal and subdural hemorrhage on CT. Subarachnoid hemorrhage detection is harder computationally because subtle SAH can appear as minimal density change in the cisterns, easily confused with beam hardening artifact. Published sensitivity numbers for SAH on non-contrast CT are generally lower than for intraparenchymal ICH, and some systems do not include SAH as a primary detection target.

Hemorrhage Volume and Triage Scoring

Beyond binary detection, ICH triage systems increasingly incorporate volumetric quantification — estimating hemorrhage volume using the ABC/2 method or fully automated 3D segmentation. Volume matters for two reasons: hemorrhage volume is a strong predictor of clinical severity and outcome, and interval volume change on serial imaging is a key decision point for neurosurgical intervention.

A 3D U-Net segmentation architecture can produce hemorrhage volume estimates with mean absolute error in the range of 1–3 ml on well-characterized intraparenchymal bleeds in published studies (Radiology, JACR). These estimates are most reliable for homogeneous intraparenchymal hemorrhage; heterogeneous bleeds with mixed density components due to coagulopathy or active bleeding are harder to segment accurately.

For triage purposes, the important output is not the precise volume number but the triage score threshold: does a 4 ml hemorrhage get escalated the same way as a 25 ml hemorrhage? Volume-weighted triage scoring — where larger hemorrhage volume produces a higher criticality index — is more clinically appropriate than binary flagging, because it allows the system to differentiate urgent-but-stable from immediately-critical findings and apply proportionate escalation.

What This Means for a Community Hospital AI Decision

The published evidence supports a genuine clinical benefit for AI-assisted ICH triage in overnight low-coverage settings. Time-to-read reductions of 45–90 minutes are plausible for community hospitals with small overnight radiology coverage, based on the mechanism and the published study findings.

The deployment decisions that maximize this benefit are: ensuring the AI system integrates with your actual worklist (not just a separate alert app that the radiologist may or may not check), configuring thresholds appropriate to your overnight CT head volume, and establishing monitoring to confirm that the triage ranking is changing radiologist reading order in the expected direction. Without the last element, you may have a working algorithm that is being behaviorally ignored — which is clinically equivalent to no algorithm at all.

For Pacslens, ICH triage is one component of a multi-indication scoring system. The ICH score is normalized to the same 0–100 criticality index as LVO, PE, and aortic dissection scores, allowing the worklist ranking to balance urgency across finding types rather than privileging one indication over others. This multi-indication approach is less well-studied than single-indication ICH detection, but it addresses the heterogeneous case mix that community hospital radiologists actually encounter overnight.