Abstract Background Pulmonary congestion is a hallmark of acute heart failure (AHF), and essential for the diagnosis, treatment, and risk stratification. We have previously shown that a low-dose chest CT has four times greater odds than chest radiographs to detect pulmonary congestion. However, implementation is limited by shortage of radiologists. We hypothesized that artificial intelligence (AI) can detect pulmonary congestion on chest CT and aimed to develop and validate an AI-algorithm to detect pulmonary congestion on chest CT. Methods We extracted all chest CT-scans ordered acutely (N= 4840 unique patients) by medical departments at X Hospital from 2016-2021. A random subset of these CT-scans (N= 2187) was used to develop an AI-algorithm to detect pulmonary congestion, using the radiology reports as reference. The remaining subset (N= 1094) was used for internal validation. We externally validated the AI-algorithm in a different cohort (X), comprising of 236 consecutive dyspnoeic patients admitted to the emergency department, all examined with a low-dose, non-contrast chest CT (LDCT). The primary outcome was pulmonary congestion adjudicated by clinical on-call radiologists with access to the medical record, and two independent, blinded research thoracic radiologists. The default AHF thresholds for the optimal cut-off, rule-in threshold and rule-out threshold was established from the internal validation. Results Of 236 patients (median age 74 years, 57% male), 62 (26%), 64 (27%) and 60 (26%) had pulmonary congestion on LDCT evaluated by the on-call radiologists and the two thoracic research radiologists, respectively. The AI-algorithm included ten discriminating features of pulmonary congestion, and the results demonstrate an AUC of 0.95 [95%CI: 0.93-0.98] to predict pulmonary congestion by the on-call radiologists. The results were confirmed by the two expert research radiologists, with an AUC of 0.91 [95%CI: 0.88-0.95] and 0.94 [95%CI: 0.90-0.97], respectively. The sensitivity and specificity of the AI-default threshold was 89% [95%CI: 78-95%] and 89% [95%CI: 83-93%], respectively. The rule-in threshold demonstrated a sensitivity of 97% [95%CI: 86-100%] and the rule-out threshold had a specificity of 96% [95%CI: 92-98%], Figure 1. Conclusion We developed and validated a novel AI-algorithm with promising potential for the rapid automatic detection of pulmonary congestion as the underlying cause of dyspnoea in consecutive acute patients.
Read full abstract