The value of multiparametric prediction scores in heart failure varies with the type of follow-up after discharge: a comparative analysis

Abstract

Aims: Multiple prediction score models have been validated to predict major adverse events in patients with heart failure. However, these scores do not include variables related to the type of follow-up. This study aimed to evaluate the impact of a protocol-based follow-up programme of patients with heart failure regarding scores accuracy for predicting hospitalizations and mortality occurring during the first year after hospital discharge.

Methods and results: Data from two heart failure populations were collected: one composed of patients included in a protocol-based follow-up programme after an index hospitalization for acute heart failure and a second one-the control group-composed of patients not included in a multidisciplinary HF management programme after discharge. For each patient, the risk of hospitalization and/or mortality within a period of 12 months after discharge was calculated using four different scores: BCN Bio-HF Calculator, COACH Risk Engine, MAGGIC Risk Calculator, and Seattle Heart Failure Model. The accuracy of each score was established using the area under the receiver operating characteristic curve (AUC), calibration graphs, and discordance calculation. AUC comparison was established by the DeLong method. The protocol-based follow-up programme group included 56 patients, and the control group, 106 patients, with no significant differences between groups (median age: 67 years vs. 68.4 years; male sex: 58% vs. 55%; median ejection fraction: 28.2% vs. 30.5%; functional class II: 60.7% vs. 56.2%, I: 30.4% vs. 31.9%; P = not significant). Hospitalization and mortality rates were significantly lower in the protocol-based follow-up programme group (21.4% vs. 54.7%; P < 0.001 and 5.4% vs. 17.9%; P < 0.001, respectively). When applied to the control group, COACH Risk Engine and BCN Bio-HF Calculator had, respectively, good (AUC: 0.835) and reasonable (AUC: 0.712) accuracy to predict hospitalization. There was a significant reduction of COACH Risk Engine accuracy (AUC: 0.572; P = 0.011) and a non-significant accuracy reduction of BCN Bio-HF Calculator (AUC: 0.536; P = 0.1) when applied to the protocol-based follow-up programme group. All scores showed good accuracy to predict 1 year mortality (AUC: 0.863, 0.87, 0.818, and 0.82, respectively) when applied to the control group. However, when applied to the protocol-based follow-up programme group, a significant predictive accuracy reduction of COACH Risk Engine, BCN Bio-HF Calculator, and MAGGIC Risk Calculator (AUC: 0.366, 0.642, and 0.277, P < 0.001, 0.002, and <0.001, respectively) was observed. Seattle Heart Failure Model had non-significant reduction in its acuity (AUC: 0.597; P = 0.24).

Conclusions: The accuracy of the aforementioned scores to predict major events in patients with heart failure is significantly reduced when they are applied to patients included in a multidisciplinary heart failure management programme.

Keywords: Heart failure; Multidisciplinary HF management programme; Prediction scores.

Figure 1

COACH Risk Engine and BCN Bio‐HF Calculator predicting heart failure hospitalizations receiver operating characteristic curves.

Figure 2

COACH Risk Engine, BCN Bio‐HF Calculator, MAGGIC Score, and Seattle HF Model predicting mortality receiver operating characteristic curves.