Deep learning-based ECG screening for chronic kidney disease


Key cohort characteristics

Cedars-Sinai Medical Center identified a total of 17,860 patients with a CKD diagnosis (7.8% of the total patient sample), of whom 7,816 had an ECG examination within 1 year of CKD diagnosis. Our primary cohort consisted of a total of 247,655 ECGs, of which 221,974 were randomly assigned to the training set (both training and validation) and 25,681 were randomly assigned to the test set. Of the ECGs in the main cohort, 74.3% did not provide serum creatinine or eGFR estimates within 30 days, and 50.7% of ECGs did not provide serum creatinine or eGFR estimates at any EHR time point. , which does not include out-of-hospital test records or paper records of clinical tests. Tests that may have been used to diagnose CKD. The mean age of the main cohort was 61.3 ± 19.7 years and 48% were female. Demographic and clinical characteristics are shown in Table 1. Demographic and clinical characteristics by age group are shown in Supplementary Table 2.

Table 1 Demographic and clinical characteristics in internal and external datasets.

Model performance in primary cohorts

Our 12-lead ECG-based model was able to discriminate all stages of CKD with an AUC of 0.767 (95% CI 0.76 to 0.773). The performance of the model was consistent across the range of CKD stages, with AUC 0.753 (0.735 to 0.770) in discriminating mild CKD, AUC 0.759 (0.750 to 0.767) in discriminating moderate-to-severe CKD, and AUC 0.783. achieved ( 0.773–0.793) in discriminating ESRD. In all cases, a negative case was defined as an ECG without CKD diagnosis.

Considering the increasing prevalence of wearable technology, especially devices containing single-lead ECG information, we aim to train additional deep learning models using information from only the single-lead ECG information and perform DLA with single-lead wearable information. simulated the performance of Using 1-lead ECG waveform data, DLA achieved an AUC of 0.744 (0.737–0.751) in detecting all stages of CKD, with sensitivity and specificity of 0.723 (0.723–0.723) and 0.643 (0.643–0.643), respectively. was.

Because early detection of CKD is important to prevent disease progression and complications in the elderly, we tested the model’s performance in young patients (<60 years old). 12-lead and 1-lead ECG-based DLA were able to detect any stage of CKD in patients younger than 60 years with AUCs of 0.843 (0.836 to 0.852) and 0.824 (0.815 to 0.832), respectively.

We also separately tested model performance among diabetic, hypertensive and elderly patients, commonly regarded as high-risk subgroups. A 12-lead based model detected CKD with AUC 0.747 (0.707-0.783) in diabetic patients, AUC 0.711 (0.696-0.725) in hypertensive patients, and AUC 0.706 (0.697-0.716) in patients aged 60 years and older. age. When the model was trained on a 12-lead ECG waveform, age, sex, diabetes, and hypertension, the model gave similar discrimination in his CKD at any stage in the heldout test set, with 0.79 (0.781–0.798) his Achieved with AUC. Detailed results of 1-lead and 12-lead ECG-based DLA performance in the holdout test set are shown in Tables 2 and 3, and AUC curves are shown in Supplementary Fig. 1.

Table 2 Performance of the 12-lead ECG-based deep learning algorithm on the internal dataset.
Table 3 Performance of one-lead ECG-based deep learning algorithms on the internal dataset.

This model performed similarly in detecting CKD in a subset population of albuminuric patients, in patients with corresponding laboratory and eGFR-proven, and in both outpatients and inpatients (Supplementary Table 3). For patients diagnosed with CKD and an estimated eGFR <60 mL/min, the AUC was 0.754 (0.737 to 0.771), and this performance was similar to that in hyperkalemia patients with AUC of 0.741 (0.698 to 0.787). The same was true in patients without hyperkalemia. Hyperkalemia with AUC 0.758 (0.747 to 0.768). The model also performed well in patients with known albuminuria, with an AUC of 0.734 (0.723–0.745) and similar performance regardless of the positive-to-negative ratio in the training set (Supplementary Table 4).

ECG features in CKD

To understand the key features associated with deep learning models that can detect CKD, we performed two sets of experiments evaluating ECG parameters important for identifying CKD. Statistically significant differences in all available ECG variables (heart rate, PR interval, P-wave duration, QRS duration, QTc interval, P-axis, R-axis, T-wave axis) between CKD stages was found to exist (Supplementary Table 5). .

We then used LIME to identify ECG segments that were specifically used to identify CKD. Supplementary Fig. 2 shows examples of LIME-highlighted ECG segments in 12-lead and 1-lead ECG waveforms taken from correctly recognized CKD patients in the holdout test set and healthy control patients. In both examples, the LIME-highlighted ECG features were primarily focused on the QRS complexes and PR intervals. In addition, QRS complexes and PR intervals in limb leads are most frequently highlighted and may indicate CKD-related electrophysiological changes.

Characteristics of the external validation cohort

The externally validated cohort consisted of 312,145 patients with a total of 896,620 ECGs. The prevalence of mild CKD was 1.2%, moderate-to-severe CKD was 3.6%, and ESRD was 0.9%. The mean age of the externally validated cohort was 56.7 ± 18.7 years and 50.4% were female. The percentage was 47.5% white, 3.6% black, 12.3% Asian, and 36.6% of other or unknown race. Demographic and clinical characteristics are shown in Table 1.

Model performance on external validation dataset

In the externally validated dataset, the 12-lead and 1-lead models performed similarly to the primary cohort. A 12-lead ECG-based model achieved an AUC of 0.709 (0.708-0.710) in discriminating all stages of CKD. A one-lead ECG-based model detected all stages of CKD with an AUC of 0.701 (0.700 to 0.702).

Consistent with the primary cohort where our model achieved higher CKD detection accuracy among younger patients, the 12-lead and 1-lead ECG-based models scored 0.784 (0.782–0.786) and We achieved an AUC of 0.777 (0.775 to 0.779). Subjects under the age of 60, respectively. Detailed results of 1-lead and 12-lead ECG-based DLA performance in the externally validated cohort are shown in Supplementary Tables 6 and 7.



Source link

Leave a Reply

Your email address will not be published. Required fields are marked *