SBRG is pleased to announce acceptance of an abstract on predictive bio-analytics for exhibition at the Conference on Innovation in Biomarkers (Journées de l’Innovation en Biologie), taking place on November 7-8, 2024, in Paris, France.
SBRG conducted research with our development and business partner, Numa Health International (France) using data obtained from a prior study with the department of cardiology at Lithuania State Medical University. Using SBRG’s proprietary biomarker drift-cluster analysis approach and NUMA’s machine learning expertise, we developed a model predicting two-year risk of major adverse cardiac events (MACE) after a first myocardial infarction. What was significant about the study was the use of factors non-specific for cardiovascular dysfunction. This would be a disruption in the medical model because it re-contextualizes the diseased organ within the global context of the whole person and the systemic adaptation response.
For more information or to explore research opportunities, feel free to reach out to our Chief Scientific Officer, Kamyar M. Hedayat, MD, at khedayat@systemsbrg.com.
Below is a translation of the abstract, which originally appears in French.
Prediction of two-year risk of MACE with a machine learning model using non-specific biomarkers and clinical variables
Kamyar M. Hedayat1,2, Ahava Cohen1, Théo Hennion1, Maël Yang1, Rima Braukyliene3 Laura Zajanckauskiene3, Martynas Jurenas3, Ramunas Unikas3, Ali Aldujeli3, Osvaldas Petrokas3, Vytautas Zabiela3, Rasa Steponaviciute3, Astra Vitkauskine3, Brigita Hedayat2, Sandrita Simonyte3, Vaiva Lesauskaite3, Jean Claude Lapraz2, Diana Zaliaduonyte3
Affiliations:
1-Numa Health International, La Rochelle, France 17000
2-Systems Biology Research Group, Chicago, IL 60626, USA.
3-Cardiology Department, Lithuanian University of Health Sciences, LT 50161 Kaunas, Lithuania.
Objective
The objective of this study is to predict two year risk of major adverse cardiac events (MACE) in survivors of a first acute myocardial infarction (AMI) using non-specific biomarkers cross-analyzed with clinical factors.
Materials and methods
In collaboration with Systems Biology Research Group (SBRG) of Chicago, USA and the Lithuanian State Medical Society (LSMU) department of cardiology, the study was run using data obtained from two prior studies conducted by SBRG and LSMU. This study consisted of 315 subjects who had complete blood count (CBC), potassium, calcium and clinical data obtained on admission to the hospital for AMI.
To study the complex interactions between these factors, we used Machine learning (ML). Biomarkers were analyzed in two ways: individually and in cluster analysis of biomarker drift, developed by SBRG. Biomarker clusters are referred to as C1, C2, etc. The data was pre-treated in order to limit noise and aberrant values, after which two methods of cross-validation were employed in the training phase (80% of subjects). A number of variables were eliminated by principal component analysis (PCA). Three ML training models were then tested: Decision Tree, Random Forest and XGBoost. Optimization of hyper parameters was effectuated for each model to improve performance.
Results
Mean age was 65, of which 64% (203/315) were male. Two year incidence of MACE was 25.4% (80/315). PCA identified 11 variables containing 90% of the information necessary for the best prediction model: CLINICAL: age, weight, BMI (body mass index) BIOMARKERS: leukocytes, MCHC (mean corpuscular hemoglobin concentration), monocytes, basophils, potassium, CLUSTERS: C3, C234, C265. The XGBoost model had the best predictive capability with a specificity of 92.8% and sensitivity of 65.1%. The Mean quadratic error (MQE) was 0.1373.
Discussion
This study demonstrates the ability to predict two year risk of MACE after AMI using biomarkers, cluster analysis and clinical factors all non-specific to AMI. This approach allows for a low-cost method of identifying patients most at risk of MACE using universally obtained admission data. We suspect that the reason non-specific factors can be predictive of a specific disorder is because they represent the transversal, multi-scalar interrelationship of underlying physiologic factors of systemic disadaptation after a life-threatening event such as myocardial infarction. The high specificity is of value in not missing those at risk. However, given the low sensitivity and risk of over-training given the size of the current data set, a future study should be conducted with a larger dataset and seeking a higher sensitivity without reducing specificity. This would enhance the robustness and generalizability of our model.
Conclusion
This was a promising proof of concept study which indicates the possibility of predicting two year risk of MACE post AMI using factors non-specific to the disorder. The study used NUMA’s machine learning to analyze SBRG’s biomarker drift cluster analysis method. This approach can decrypt subtle risk factors which are apparent only when studying the qualitative interactions between them.