Our study demonstrated that examining arterial pulsation and carotid PWs alone can identify and help with the risk stratification of patients with HF in clinical practice. The DNN model had an excellent classification ability to distinguish between HF and non-HF PWs after training with PW data. Conventional machine learning exhibited comparable abilities, especially SVM, even matching the DNN model in some situations. However, with an increasing amount of training data, the performance of the DNN model improved substantially, whereas that of the SVM stagnated at a certain level. These findings are in line with the concept that deep learning performs even better with larger datasets versus machine learning [48]. If a larger dataset can be collected to enhance the DNN model, the deep learning model is expected to be more helpful than machine learning in clinical practice.
The conventional noninvasive assessment of pulse waves, such as carotid-femoral pulse wave velocity and augmentation index, has been suggested as a useful clinical assessment of arterial aging and predictor of cardiovascular events from the past [49, 50] to the present [51, 52]. Nevertheless, aside from some investigations discussing the prognostic value of pulse pressure [53] and pulse wave velocity [54], advanced evaluations of the association between PWA and prognosis in patients with HF are scarce. As such, our study represents pioneering efforts as a novel approach obtained from the information provided by PWs to help improve the clinical care of patients with HF.
Regarding the important contributing features of PW to the HF score, the machine learning findings agreed with those obtained from our DNN model in some portions of the PW; nevertheless, the top 20 features between them seem quite different (Supplementary Table 3), and the Spearman correlation indicated that SHAP values were significantly correlated with neither the beta coefficients of LR nor those of SVM. This implies that machine learning and deep learning analyze the characteristics of PW in divergent ways, which may be the reason behind the DNN model surpassing machine learning under certain conditions. On the other hand, SHAP values suggested that the systolic portion of the PW had high negative importance (non-HF) and that the dicrotic notch along with the diastolic portion affected the positive prediction (HF) (Supplementary Fig. 7E). Given the unrevealed information about PW, the DNN model should have detected some important waveform features from the interactions between the upslope and downslope of PW. Therefore, future mechanistic studies are needed to reveal the relationships of these identified portions with HF and provide deeper insight [55–57].
In the Cox regression analysis, the HF score was a significant predictor in the multivariate analysis (Table 3), which supports the valuable role of PWs in the management of HF patients. After removal of the HF score from the Cox regression, the C-index decreased and the AIC increased, supporting our conclusion that the HF score helps improve the prognostic accuracy in patients with HF. Other clinical data, such as age, serum sodium level, NT-pro-BNP, Hgb, and history of PCI, were risk factors identified by the multivariate model. The identification of these established risk factors suggests the internal validity of our data.
Compared with other risk stratification models, we found that HF score model demonstrated a superior prognostic value and an even better than conventional risk prediction model according to the C-index, which was 0.71 for the HF score model and 0.63, 0.68, and 0.69 for the ADHERE, GWTG-HF, and MAGGIC, respectively. Regarding our six significant variables, sodium serum level is also a predictor in the GWTG-HF risk score, while age is commonly recognized as a predictor by the GWTG-HF and the MAGGIC. PCI is also a traditional risk factor for HF, and it suggests the comorbidity of coronary artery disease, which is associated with an increased mortality rate [58]. Finally, NT-pro-BNP is a remarkable risk predictor of mortality in patients with HF [59–61], which was confirmed here. Remarkably, blood urea nitrogen (BUN), serum creatinine (Cr), and pulse rate (PR) were statistically significant predictors in our univariate Cox regression and predictors in all three conventional models, but none were statistically significant in the multivariate model.
The novel findings of our study on the utility of PWs require external validation. However, our results will lead to new diagnostic opportunities in cardiovascular medicine by providing insight on the use of AI technology and its possible improvement. Further studies should be conducted to relate PW and PW-like waveforms, such as photoplethysmography, to more cardiovascular diseases and, in turn, uncover its corresponding pathophysiology.
Limitations
PWs are limited to data originally collected from other cohorts [25], and future prospective cohort studies are required to validate our study findings. Due to the “black box problem” of deep learning [62], there is no certain way to explain the output from DNN model; this is also why used only SHAP to identify important PW portions. Applanation tonometry was used to obtain the PWs. Whether PWs acquired by other techniques, such as photoplethysmography or oscillometry, can result in comparable findings requires further confirmation.