Prioritizing ICH Using AI
A simulation study on pros and cons of utilizing AI for prioritization.
Derek Zhi, Feng Gao, Xiaoxiao Liu
Among many types of critical findings in radiology, intracranial hemorrhage (later referred to as ICH) requires special attention. ICH is a type of bleeding that happens inside the skull. Every year, 67000 patients in the US suffer from ICH. Though affecting a small percentage of the population, ICH has very deadly impact on individuals – the 1-month mortality rate is 35% to 52% with nearly half of the resulting mortality occurs in the first 24 hours of onset .
ICH condition requires immediate medical attention after diagnosis to mitigate the ensuing damage. Currently, medical diagnosis of ICH is based on non-contrast head computed tomography (CT). While it’s not particularly time-consuming to interpret one non-contrast head CT study which takes roughly 7 minutes, it could take a while before one study is interpreted. In order to understand why that happens, it’s necessary to understand the clinical workflow of a radiologist.
The standard of care workflow goes like the following. Upon data acquisition, the medical images are sent to integrated healthcare system’s picture archiving and communication system (PACS). An ordering physician inspects the incoming studies in PACS and assigns the priority to the study. The priority is usually determined by patient status (outpatient vs. inpatient). A “routine” priority is usually assigned to outpatient and a “stat” priority to inpatient or patients with emergency. Stat studies are typically interpreted within an hour while routine studies can take much longer. The studies then enter the worklist which is comprised of two queues – a stat queue and a routine queue. Within each queue, the studies are processed in a first-in-first-out (FIFO) order. Furthermore, studies in the stat queue are always processed before the ones in the routine queue. An illustration is the following.
To overcome this challenge, we explore an AI-assisted prioritization algorithm (later referred to as the Algorithm) that detects ICH from non-contrast head CT studies. The Algorithm alters the standard of care workflow to the following.
Monte Carlo Simulation
Primary end point: turnaround time (TAT) comparison between the standard of care and AI-assisted ICH prioritization. TAT is the time duration from a study showing up on the worklist to radiologists finishing the reporting of this study.
Model parameters as random variables that follow truncated normal distributions :
The following parameter values are used to represent different clinical settings where the composition of the STAT patients, the head CTs, and the prevalence of ICH are used in the simulation.
- The ratio of STAT and ROUTINE studies:
- 20% stat, 80% routine
- 40% stat, 60% routine
- The ratio of non-contrast head CTs among all studies:
- 5% non-contrast head CT
- 10% non-contrast head CT
- 20% non-contrast head CT
- The rate of occurrence of ICH in ROUTINE studies:
- The rate of occurrence of ICH in STAT studies:
AI Model performance:
We assume the AI algorithm has 90% precision and 90% recall.
AI-prioritization Compared to Standard of Care Prioritization Scheme
Based on our simulation results in varying clinical senarios, compared to the standard of care prioritization, the new workflow with AI capability reduces the TAT significantly (from 35% to 96%) for studies containing ICH findings. The effects of potential delays on other critical conditioned patients and the AI-missed ICH patients (compared to the standard of care) are from 0.3% to 7.2%. When the STAT percentage (out of the total imaging studies) if small (up to 20%), the caused delays are negligible (up to 3.2%).
In a typical mid-sized radiology practice (where the head CT image studies are 10% of total incoming imaging studies and the percentage of STAT ordering is 20%), our simulation shows that the TAT for the majority of ICH patients will be saved by 94%, with 3.2% potential delay on other critical condition patients.
 Dora, J. M., Torres, F. S., Gerchman, M., & Fogliatto, F. S. (2016). Development of a local relative value unit to measure radiologists’ computed tomography reporting workload. Journal of medical imaging and radiation oncology, 60(6), 714-719.
 Shinagare, A. B., Ip, I. K., Abbett, S. K., Hanson, R., Seltzer, S. E., & Khorasani, R. (2014). Inpatient imaging utilization: trends of the past decade. American Journal of Roentgenology, 202(3), W277-W283.
 Raja, A. S., Ip, I. K., Sodickson, A. D., Walls, R. M., Seltzer, S. E., Kosowsky, J. M., & Khorasani, R. (2014). Radiology utilization in the emergency department: trends of the past 2 decades. American Journal of Roentgenology, 203(2), 355-360.