This is a prospective observational study with two independent cohorts. One of the cohorts (DEXA) included 96 consecutive patients admitted at the Cardiology Department of Clinical Hospital of Santiago de Compostela between 2014 and 2015. The other cohort (SEC-ROVI) included 122 patients admitted at the Cardiology Department of Clinical Hospital of Santiago de Compostela and Pontevedra between 2018 and 2019. Heart failure diagnostic was made according to the recommendations of the European Society of Cardiology (Ponikowski et al., 2016).

The only exclusion criterion was the previous HF history. The database collected demographic, clinical (electrocardiogram and echocardiogram parameters within 24 h after admission) and laboratory analysis [haemogram, basic biochemistry and coagulation rate, lipid and glucose profile, as well as specialized parameters such as levels of electrolytes and pro-brain natriuretic peptide (NT-proBNP)], registered after admission.

The study complies with the Declaration of Helsinki and was approved by the Clinical Research Ethics Committee of Galicia. All patients provided informed consent.

Blood samples were obtained at discharge and were centrifuged at 1,800 × g for 15 min. Isolated plasma was stored at –80°C until use. Orosomucoid levels from the DEXA cohort were analyzed by ELISA with a detection limit of 3,91 ng/mL (GenWay Biotech Inc.) (Agra et al., 2017) and from the SEC-ROVI cohort by ELISA with a detection limit of 59 ng/mL (SEA816Hu, Cloud Clone). Omentin levels were measured by ELISA with a detection limit of 6.5 pg/mL (SEA933Hu, Cloud Clone Corp, Houston, United States) following the protocol of the manufacturer.

NT-proBNP levels were determined at admission by the clinical laboratory according to the manufacturers’ protocol with a Roche reagent pack on an Elecsys 2010 immunoanalyzer (Roche Diagnostics) and European guidelines (McDonagh et al., 2021).

The primary endpoint was death from any cause and/or HF rehospitalization. Follow-up information was recorded from medical history. The mean time of follow-up was 317 (3–575) days.

Continuous variables were represented by the mean and standard deviation (SD) or median and interquartile range if the distribution of the variables were not normal. Differences between the patients of each project were performed by ANOVA test or a Kruskall-Wallis test in the absence of normality. The categorical variables were represented with the absolute and relative frequencies and the differences between projects were evaluated using the χ²-test, or with the χ²-test with permutation if the expected frequency in any of the cells is less than 5. Differences in variables between patients who suffered or did not suffer an event (rehospitalization for HF and/or death) were determined by ANOVA test, or Kruskall-Wallis test, according to the normality of continuous variables. The differences on categorical variables were evaluated using the χ²-test, or with the χ²-test with permutation if the expected frequency in any of the cells was less than 5. For both continuous and categorical variables, the odds ratio (OR) was obtained by maximum likelihood estimation (MLE), and its 95% confidence interval (normal approximation) was represented. All survival analyzes were carried out using the Kaplan-Meier method. In addition, the differences in survival were evaluated based on each of the three categorized markers (according to the optimal cut-off point), using a Log-rank test. The optimal cut-off point for the markers was estimated with a methodology based on correlation with survival time: Maximally selected rank statistics. The cut-off point was the one that provides the most significant split based on the standardized log-rank test.

For the predictive models, several steps were performed:

Regarding sample size estimation since the principal concern in clinical predictive modeling is the reduction of overfitting and precise estimation of parameters, we followed the sample size estimation method proposed by Riley et al. (2020). For estimating sample size with this methodology three criteria must be met: (a) small optimism (overfitting) in the estimation of model parameters which is reached by a shrinkage or penalization factor of 0.9 (the predictions in new patients will move away from the extremes and toward the mean), (b) reduce the difference between apparent and adjusted Nagelkerke’s \(R^2\) to a small value of 0.05, and (c) increase precision (\(\pm\) 0.05) in the estimation of the average risk in a selected time point in the target population. The required sample size corresponds with the one with the highest value among the three criteria. Required information is obtained from previous studies.

Furthermore, with the aim of minimizing problems regarding the overfitting due to a small sample size, regularization methods for selecting candidate predictors were used as well as internal validation to reduce the optimism in estimation.

The statistical analysis was performed by R software. The packages mice (Rubin, 1978; van Buuren, 2007; Groothuis-oudshoorn, 2011; Galbraith, 2012), miceadds (Robitzsch and Grund, 2020), and Hmisc (Harrell and Frank, 2018) were used for imputation and pooling; timeROC (Blanche et al., 2013) for obtaining time-dependent ROC curves; glmpath (Park and Hastie, 2018) for model selection, and rms (Harrell and Frank, 2018) for model selection, validation, calibration, and prognostic models presentation.

After missing some clinical values in one of the two cohorts from 17 patients, we included the main clinical characteristics of 201 patients for their comparison (Supplementary Table 1). The mean age was 71.0 (61.0;79.0) years old with a BMI of 29.0 (26.8;33.6) kg/m². Of all patients, 65.7% were men, 39.6% drank alcohol, 22.2% were smokers, 35.3% were diabetic, 68.2% were hypertensive, 47% had edemas at admission, and 12.9%, chronic obstructive pulmonary disease (COPD). The main differences between the two cohorts (DEXA and SEC-ROVI) were alcohol intake, BMI, percentage of patients with edemas at admission, hepatomegaly, and previous stroke.

Orosomucoid levels were measured in 160 patients stratified in 3 quantiles (Q1, the lowest levels, Q2 and Q3, the highest levels). Those patients with highest orosomucoid levels (Q3 or Q2) had higher glucose levels 130 (98.2;177) or 163 (121;243) mg/dL, respectively, than those patients in Q1 122 (106;183], p = 0.008. The highest levels of creatinine were described in Q3 patients 1.21 (0.86;1.47) vs. 0.95 (0.77;1.10) in Q1 or 0.87 (0.75;1.02) mg/dL in Q2, p = 0.001. The percentage of patients with an event was higher in the Q3 group 16 (31.4%) vs. 6 (10.3%) in Q1 or 9 (15.0%) in Q2, p = 0.013 (Table 1). Regarding omentin, which was measured in 155 patients, a higher percentage of smokers and alcohol intake was represented on the Q1 group (lowest omentin levels) (Table 2). The ischemic etiology was represented in a similar percentage among groups of orosomucoid or omentin quantiles. The orosomucoid Q1, Q2, and Q3 were represented by 16, 19, and 25%; p = 0.458 and the omentin Q1, Q2, and Q3 by 18, 24, and 20%; p = 0.774, respectively.

The correlation among markers showed that neither orosomucoid nor omentin showed statistically significant correlations with any clinical variable (Supplementary Figure 1).

We performed the imputation of the lost values using the MICE algorithm (Multiple imputation by chained equations). Confidence intervals obtained by overimputation showed a great degree of coverage for all the variables (for all the imputed variables, the overimputed confidence interval included the observed values in more than 90% of the cases). The distribution of imputed and original values was very close. Afterward, the median was defined as the optimal cut-off point, threshold to consider high or low values of the markers, for orosomucoid and omentin. The omentin and orosomucoid cut-off values were 6.637 and 1.409, respectively. Then, the following categories were generated: (a) DownUp: Orosomucoid values below optimal and omentin values above optimal, (b) Equal: both proteins were simultaneously above the optimum, or below it, (c) UpDown: orosomucoid values above optimal and omentin values? below optimal. The main differences among groups were the higher percentage of Chronic obstructive pulmonary disease (COPD), creatinine and potassium levels at admission in the UpDown group (Table 3).

During the follow-up, 16% of patients were rehospitalized for HF and/or death. These patients were older, mostly hypertensive, with high NT-proBNP and high creatinine levels at admission with respect to those without events (Supplementary Table 2). Patients in the UpDown group had the worse prognosis at 90, 180, and 365 days. The Kaplan-Meier curves showed that these differences were statistically significant (p < 0.05) (Figure 1).

After testing the number of events in the NT-proBNP, orosomucoid, and omentin categorized groups by quantiles and the combined orosomucoid and omentin group regarding cut-off values (Up or down), we observed that all the events in the DownUp group, low orosomucoid, and high omentin levels, were recorded 6 months after discharge (Supplementary Figure 2). Contrary, the UpDown group had events in the following 30 days after discharge.

Pathologies were grouped in a Score, assigning 3 points if the patient had hypertension, 2 points if they have COPD, 1.2 points if they have hyperlipidemia, and 1 point to the remaining pathologies. For the treatments, the variable Pharmacological-score was generated, assigning 4 points if the patient takes nitrates, 3 points if the patient takes amiodarone, 2 points if the patient takes acetylsalicylic acid, and 1 point to the other drugs. Three different methodologies, the Lasso method and Backward Step-down deletion with and without bootstrapped validation (200 bootstrap samples) with AIC as a stopping rule, were performed for selecting the variables for the best predictive model. Among the most frequently selected covariates were LVEF, K⁺, Creatinine, BMI, and pharmacological and pathology scores. NT-ProBNP was selected in every model. With regard to Omentin and Orosomucoid, their selection was increased when their interaction with any of the scores was also included (Supplementary Table 3). After testing the predictive models that include the same clinical variables of BMI, pathology and pharmacological scores, LVEF and K^+, and the biomarkers, we observe that the inclusion of NT-proBNP determined a c-index = 0.8016 (optimism corrected C-index = 0.798) (Central Illustration), the inclusion of orosomucoid depicted a c-index = 0.8036 (optimism corrected C-index = 0.788) (Figure 2), the inclusion of ometin a c-index = 0.7953 (optimism corrected C-index = 0.772) and the combination of categorized orosomucoid and omentin (UpDown, Equal or DownUp) (OROME) had similar c-index = 0.7889 (optimism corrected C-index = 0.778) (Figure 2). But the inclusion of NT-proBNP and the combined orosomucoid and omentin (OROME) showed the best c-index = 0.8488 (Figure 2). These results indicated the improvement of the predictive model with all three biomarkers (Figure 2).

The AUC were sustained in the time (Supplementary Figure 3A) with model OROME + NT-proPNB outperformed model NT proBNP (Supplementary Figure 3B). Models including only combination of orosumucoid and omentin (OROME) or orosomucoid performed as well as NT proBNP model (Supplementary Figures 3C,D).

The additional value of the new biomarker according NRI is 31.2% (95%CI = 0.01–54.8%, p-value = 0.05), with and IDI of 4.9% (95% CI = 0.1–12.1%, p-value = 0.04).

Calibration plots showed very good concordance between predicted and observed survival probabilities for the selected model (Supplementary Figure 4).

Nomogram of the finally selected model (including BMI, pathology and pharmacological scores, LVEF, K⁺, NT-proBNP, and OROME) is presented in Figure 3. It shows predicted survival probabilities at several follow-up times. The total points are obtained by adding the individual points of each predictor. These points are obtained by drawing a vertical straight line from the predictor value to Points (i.e., 42.5 points are assigned to a concentration of NT-ProBNP of 15,000 ng/mL).

The number of patients included is small but is comparable to other studies that assessed similar objectives and outcomes and the results are statistically significant and we think that this issue will have future clinical relevance. The adjustment for other clinical parameters and biomarkers may influence our results.