Real-World Data and Machine Learning to Predict Cardiac Amyloidosis
Elena García-García1, Gracia María González-Romero1, Encarna M Martín-Pérez2
1Fundación San Juan de Dios, Centro CC de la Salud San Rafael, Universidad Nebrija, 28036 Madrid, Spain.
Insights
This study introduces a statistical learning algorithm to detect cardiac amyloidosis, also known as "stiff heart syndrome." The algorithm effectively identifies this rare heart condition using clinical records, aiding in early diagnosis.
Area of Science:
- Cardiology
- Medical Informatics
- Machine Learning
Background:
- Cardiac amyloidosis, or "stiff heart syndrome," is a rare restrictive cardiomyopathy characterized by amyloid deposits in the heart muscle.
- Delayed diagnosis is common due to the condition's rarity and diagnostic challenges, often leading to a poor prognosis.
Purpose of the Study:
- To analyze the characteristics of cardiac amyloidosis.
- To propose and validate a statistical learning algorithm for detecting the disease using clinical records.
Main Methods:
- Utilized hospitalization clinical records (medical and nursing) for algorithm training and learning.
- Treated patient admission and discharge data as vectors, addressing high dimensionality and class imbalance due to low disease prevalence.
Main Results:
- Demonstrated the feasibility of learning from limited clinical data for rare diseases like cardiac amyloidosis.
- The algorithm successfully detected the disease when applied to general and heart failure patient data, identifying disease episodes through data vector profiling.
Conclusions:
- The developed statistical learning algorithm shows potential for screening specific populations to improve early detection of cardiac amyloidosis.
- This predictive technique can aid in identifying patients with this challenging heart condition.
Abstract:
(1) Background: Cardiac amyloidosis or "stiff heart syndrome" is a rare condition that occurs when amyloid deposits occupy the heart muscle. Many patients suffer from it and fail to receive a timely diagnosis mainly because the disease is a rare form of restrictive cardiomyopathy that is difficult to diagnose, often associated with a poor prognosis. This research analyses the characteristics of this pathology and proposes a statistical learning algorithm that helps to detect the disease. (2) Methods: The hospitalization clinical (medical and nursing ones) records used for this study are the basis of the learning and training techniques of the algorithm. The approach consisted of using the information generated by the patients in each admission and discharge episode and treating it as data vectors to facilitate their aggregation. The large volume of clinical histories implied a high dimensionality of the data, and the lack of diagnosis led to a severe class imbalance caused by the low prevalence of the disease. (3) Results: Although there are few patients with amyloidosis in this study, the proposed approach demonstrates that it is possible to learn from clinical records despite the lack of data. In the validation phase, the algorithm first acted on data from the general study population. It then was applied to a sample of patients diagnosed with heart failure. The results revealed that the algorithm detects disease when data vectors profile each disease episode. (4) Conclusions: The prediction levels showed that this technique could be useful in screening processes on a specific population to detect the disease.
