# Um Novo Escore de Risco Baseado em Aprendizado de Máquina (Machine Learning) em Pacientes com Insuficiência Cardíaca Aguda: O Escore ML-HF

## TL;DR

A new machine learning-based risk score for predicting in-hospital death in acute heart failure patients outperforms traditional scores.

## Contribution

A machine learning model (ML-HF) was developed and validated to predict mortality in acute heart failure patients more accurately than traditional scores.

## Key findings

- The ML-HF score showed better discrimination (AUC = 0.722) than traditional scores like GWTG-HF and ADHERE.
- Key predictors included WHOQOL-BREF physical health, serum sodium, urea, creatinine, and systolic blood pressure.
- The model was validated on 30% of data and showed adequate calibration.

## Abstract

Os escores de avaliação prognóstica convencionais muitas vezes não têm desempenho suficiente para prever a mortalidade em pacientes com insuficiência cardíaca aguda (ICA).

Desenvolver e validar um escore prognóstico baseado em aprendizado de máquina (
Machine Learning - escore ML-HF)
para prever morte hospitalar em pacientes com ICA e comparar seu desempenho com os principais escores tradicionais.

Pacientes admitidos por ICA em hospitais brasileiros do “Programa de Boas Práticas em Cardiologia” de 2016 a 2022 foram incluídos. Dados clínicos, resultados laboratoriais e o questionário de Qualidade de Vida da Organização Mundial da Saúde (
World Health Organization Quality of Life)
WHOQOL-Bref foram coletados na admissão hospitalar. O desfecho foi óbito hospitalar. O modelo foi treinado usando 70% das admissões (conjunto de treinamento) e validado com os 30% restantes (conjunto de teste). A área sob a curva ROC (AUC) do escore ML-HF foi comparada com a AUC dos escores tradicionais
Acute Decompensated Heart Failure National Registry
(ADHERE) e
Get With the Guidelines–Heart Failure
(GWTG-HF). O nível de significância foi de p < 0,05.

Foram incluídos mil cento e cinquenta e sete pacientes hospitalizados por ICA. As cinco variáveis mais importantes do escore ML-HF foram: Qualidade do Domínio de Saúde Física (WHOQOL-BREF), sódio sérico, ureia sérica, creatinina sérica e pressão arterial sistólica na admissão hospitalar. No conjunto de teste, o escore ML-HF apresentou calibração de modelo adequada (valor de p do teste de Hosmer-Lemeshow = 0,056) e discriminação (AUC = 0,722 [IC95%, 0,661-0,783]), que foi superior aos escores GWTG-HF (AUC = 0,616 [IC95%, 0,529-0,702; p = 0,014]) e ADHERE (AUC = 0,601 [IC95%, 0,511-0,691; p = 0,006]).

Desenvolvemos e validamos um escore usando aprendizado de máquina para prever morte hospitalar em pacientes com ICA, que superou os escores tradicionais.

Figura Central:Um Novo Escore de Risco Baseado em Aprendizado de Máquina (Machine Learning) em Pacientes com Insuficiência Cardíaca Aguda: O Escore ML-HF

Conventional prognostic assessment scores often fall short in performance to predict mortality in patients with acute heart failure (AHF).

To develop and validate a machine learning-based prognostic score to predict in-hospital death in patients with AHF and compare its performance with the main traditional scores.

Patients admitted for AHF in Brazilian hospitals from the “Best Practices in Cardiology Program” from 2016 - 2022 were included. Clinical data, laboratory results, and the World Health Organization Quality of Life (WHOQOL-Bref) at hospital admission were collected. The outcome was in-hospital death. The model was trained using 70% of admissions (training set) and validated with the remaining 30% (test set). The ML-HF score area under the ROC curve (AUC) was compared with the AUC of the traditional scores Acute Decompensated Heart Failure National Registry (ADHERE) and Get With the Guidelines–Heart Failure (GWTG-HF). The level of significance was p<0.05.

One thousand and one hundred fifty-seven patients hospitalized for AHF were included. The five most important variables of the ML-HF score were: Physical Health Domain Quality (WHOQOL-BREF), serum sodium, serum urea, serum creatinine, and systolic blood pressure at hospital admission. In the test set, the ML-HF score showed an adequate model calibration (Hosmer-Lemeshow test p value=0.056) and discrimination (AUC=0.722 [CI95%, 0.661-0.783]), which was superior to GWTG-HF (AUC=0.616 [CI95%, 0.529-0.702; p=0.014]) and the ADHERE (AUC=0.601 [CI95%, 0.511-0.691; p=0.006]) scores.

We developed and validated a score using machine learning to predict in-hospital death in patients with AHF, which outperformed the traditional scores.

Central Illustration:A New Risk Score Based on Machine Learning in Patients with Acute Heart Failure: The ML-HF Score