Improved risk prediction using machine learning on perfusion CMR and 10-year registry data

Presenting Author(s)

• 

  Mathias Permlid, MSc

  Researcher
  Lund University
  Lund, Skane Lan, Sweden

Primary Author(s)

• 

  Mathias Permlid, MSc

  Researcher
  Lund University
  Lund, Skane Lan, Sweden

Co-Author(s)

• MK

  Mattias Karlsson, PhD

  Postdoctoral researcher
  Lund University, Sweden
• SA

  Shahnaz Akil, MSc

  Technologist, PhD
  Lund university and Skåne University Hospital, Lund
  Lund, Skane Lan, Sweden
• 

  Ellen Ostenfeld, MD, PhD

  MD, PhD, Associate professor
  Lund University and Skåne University Hospital, Lund, Sweden
  Lund, Skane Lan, Sweden
• BK

  Barbro Kjellström, PhD

  PhD
  Lund University and Skåne University Hospital, Lund, Sweden, Sweden
• 

  Peter Kellman, PhD

  Senior Scientist
  National Institutes of Health
  Bethesda, Maryland, United States
• HA

  Håkan Arheden, MD, PhD

  MD, PhD, Professor
  Lund University and Skåne University Hospital, Lund, Sweden
  Lund, Sweden
• 

  Morten Kraen, MD

  physician
  Lund University
  Malmo, Skane Lan, Sweden
• HE

  Henrik Engblom, MD, PhD

  Professor, MD
  Clinical Physiology, Department of Clinical Sciences Lund, Lund University, Skåne University Hospital, Lund, Sweden
  Lund, Skane Lan, Sweden

Background: Accurate risk prediction for coronary heart disease is essential for guiding patient therapy and follow-up. While perfusion CMR is known for its prognostic value, it is not well studied whether machine learning could be useful in harvesting the prognostic potential of other known risk factors and comorbidities. This study aimed to quantify the combined incremental value of perfusion CMR metrics with 10 years of historical registry data of prior diagnoses and medical procedures in predicting major adverse cardiac events (MACE) in patients with suspected ischemic heart disease.

Methods: In a clinical database, a total of 2,239 patients referred for adenosine perfusion CMR were identified; 80 were excluded due to missing data, leaving 2,159 for analysis. All CMR exams were performed on a 1.5T scanner using a single-bolus, dual-sequence technique. From the national health registry, diagnoses and procedures related to cardiac health were extracted for 10 years preceding the CMR examination. Demographic data on age, sex, and BMI were collected at the time of CMR. The CMR variables consisted of LVEF, indexed LVEDV/LVESV/LV mass, and presence or absence of infarct and ischemia. MACE was defined as cardiovascular death, myocardial infarction, unstable angina, or coronary intervention. For the prediction of MACE, Random Forest (RF) models were trained. The ROC area under the curve (AUC) was used to summarize and discriminate the different models. Statistical significance was tested with the DeLong test for correlated ROC curves.

Results: In the study population (n = 2159), the mean age was 65.4 ± 12.2 years, 44% were women, and the mean BMI was 27.6 ± 5.1 kg/m². The mean follow-up time was 2.5 ±1.0 years, and MACE occurred in 9.1%. Using the registry and demographic data to train the RF, an AUC of 0.735 was achieved (Figure 1). Adding the CMR metrics significantly increased the AUC to 0.818; Δ = 0.083 (p < 0.0001). Notably, the model including registry data, demographics, and CMR metrics also outperformed the model with CMR and demographics data alone (AUC 0.786; Δ = 0.032; p< 0.0001). Permutation importance analysis identified stress-induced ischemia as the dominant predictor (Figure 2).

Conclusion: In patients with suspected ischemic heart disease, perfusion CMR findings were of major prognostic value. Moreover, machine learning methods incorporating historical data on comorbidity significantly improved the risk prediction.

Figure1. ROC curves for pre-CMR (red) vs post-CMR (blue) (Δ = 0.083), including individual feature sets.
Figure2. Top ten features with the highest permutation importance.