1 Introduction
2 Outline of the paper
3 Previous systematic literature reviews on ML and HF dataset
4 Methods
4.1 Identification of studies and literature searches
4.2 Inclusion and exclusion criteria
5 Results
5.1 Common HF problems addressed by ML
Outcomes of studies | N/% | Citations to studies |
---|---|---|
Detection of HF onset | 25 (31%) | |
Mortality prediction | 20 (26%) | |
Prediction of readmission to hospital | 18 (21%) | |
Classification of HF (according to NYHA class or aetiology) or clustering | 11 (13%) | |
Other: * | 5 (6%) |
5.2 General study characteristics: variables, source and size of datasets used.
Discrete data | Continuous data |
---|---|
Demographics | Physical examination: |
Age / Sex / Gender: Female/ Male | Pulse Rate (beats per minute) |
Race: White, African American, Hispanic, American Indian,Native Asian | Respiratory Rate (breath per minute) |
Medicare Insurance / Medi-Cal Insurance | Systolic Pressure in mmHg |
Contacts with Healthcare/Care management: | Body Mass Index (BMI) |
Discharge to Skilled Nursing Facility | Body Surface Area (BSA) |
Missed Clinic Visits in Prior Year | Blood tests |
ED and O/P Visits in Prior Year | Biochemistry: |
Admission in Previous Year | Serum B- Natriuretic Peptide (pg/mL) |
In-admission Telemetry Monitoring | Glucose (mg/dL), |
Clinical data | Fasting blood glucose >120 mg/dl |
Symptoms: | Serum Creatinine (mg/dL) |
Types of chest pain, Breathlessness as per NYHA class | Urea (mg/dL) |
Past Medical History: | Serum Albumin (g/dL) |
Ischemic Heart Disease | Cholesterol level |
Previous Myocardial Infarction | Serum Sodium and Potassium (mEq/L) |
Previous Heart Failure | Haematology: |
Type of Cardiomyopathy | Haemoglobin (g/dL), Haematocrit (%) |
Coronary Artery Disease | Additional tests |
Valvular Heart Disease | ECG (recorded at rest) features |
Arrhythmias | heart rhythm: sinus rhythm atrial fibrillation |
Cerebrovascular Disease/Stroke/TIA | ORS width - broad or narrow |
Vascular/Circulatory Disease | Exercise Stress Test (EST): |
Diabetes type I and II | MPHR - maximum predicted heart rate |
Renal Disease or ESRD on Dialysis | EST induced angina, ST segment depression, downslope of ST segment or upslope of ST segment |
Chronic Lung Disease/COPD/Asthma | ECHO features |
Metastatic Cancer of solid organ or Acute Leukemia | LVEF in % (left ventricular Ejection Fraction) |
Severe haematological disorder | Right ventricular systolic pressure |
Liver Disease | Pulmonary artery mean pressure |
Mental Disorder(s) | Chest XRay features: |
Medication History | Lung fields |
Social History: Alcohol Abuse, Drug Abuse, Protein Caloric Malnutrition, Functional Disabilities | Cardiomegaly |
Study outcome | Median number of variables | Most common variables |
---|---|---|
Detection of HF | 14 (range: 13 - 1823) | Age, presence of: diabetes, hypertension |
HF mortality prediction | 45 (range: 8 - 1302) | LVEF, comorbidities, age, renal function tests (creatinine, urea) |
HF classification | 55 (range: 11 - 400) | Hypertension, age, gender, coronary artery disease, blood tests, renal function tests |
Prediction of HF readmission | 56 (range: 16 - 4205) | Age, blood tests, comorbidities |
5.3 Dimensionality of datasets
5.4 Handling of missing data
5.5 Overview of algorithms
ML alghorithms | N (%) number of (models (% out of 81 studies ) |
---|---|
Supervised ML methods | |
Logistic regression (LR) (15) (including Boosted LR, Regularised LR, Knowledge Driven Scalable Orthogonal Regression, Spike-and-slab regression, Multivariable regression, Stepwise LR, Ensemble LR) | 22 (26%) |
Decision Tree (including Decision Tree ID3 (10), Boosted Decision Tree (2), Boosted Regression Tree (3) | 15 (18%) |
Random Forest | 14 (17%) |
Support Vector Machine (SVM) | 12 (14%) |
Naive Bayes (NB) (including NB (5), Tree Augmented NB (2), Gaussian NB (1)) | 8 (9%) |
Deep learning | |
Neural Networks (NN) (including Recurrent NN (7), Convolutional NN (2), Deep NN (1)) | 10 (12%) |
Unsupervised ML | |
Clustering Methods (including k nearest neighbour (6), k-means clustering (3), hierarchical clustering (1)) | 9 (11%) |
Selection Operator Models | 1 |
Feature Rankin Analysis | 1 |