PaperlayerView exact revision on arXiv

Source-linked AI summary

Machine learning based disease diagnosis: A comprehensive review

Md Manjurul Ahsan, Zahed Siddique

arXiv:2112.15538v1cs.LG

TL;DR

Disease diagnosis is difficult because disease mechanisms and symptoms are complex, while symptoms may be ambiguous and healthcare systems may lack sufficient professionals. The paper reviews ML and DL applications for early disease identification, combining bibliometric analysis of Scopus and WOS publications with a synthesis of algorithms, diseases, data, applications, and evaluation metrics. Across the reviewed literature, CNN is reported as the most dominant algorithm, while the review’s coverage is constrained by its database selection and level of methodological detail.

  • Problem

    Ambiguous symptoms, rare diseases, overlooked conditions, and insufficient health professionals create challenges for effective disease diagnosis.

  • Method

    The review analyzes Scopus and WOS literature bibliometrically and synthesizes MLBDD approaches by algorithm, disease type, data type, application, and evaluation metrics.

  • Results

    CNN is reported as the most dominant MLBDD algorithm, with SVM, RF, DT, transfer learning, and ensemble techniques also widely represented.

  • Takeaways & Limitations

    The review aims to guide novice and expert researchers and practitioners in selecting ML and DL methods for disease diagnosis.

  • Takeaways & Limitations

    The review highlights ML and DL algorithms, datasets, disease classifications, and evaluation metrics but does not examine the detailed ML processes reported in the reference literature.

Abstract

from arXiv · show

Globally, there is a substantial unmet need to diagnose various diseases effectively. The complexity of the different disease mechanisms and underlying symptoms of the patient population presents massive challenges to developing the early diagnosis tool and effective treatment. Machine Learning (ML), an area of Artificial Intelligence (AI), enables researchers, physicians, and patients to solve some of these issues. Based on relevant research, this review explains how Machine Learning (ML) and Deep Learning (DL) are being used to help in the early identification of numerous diseases. To begin, a bibliometric study of the publication is given using data from the Scopus and Web of Science (WOS) databases. The bibliometric study of 1216 publications was undertaken to determine the most prolific authors, nations, organizations, and most cited articles. The review then summarizes the most recent trends and approaches in Machine Learning-based Disease Diagnosis (MLBDD), considering the following factors: algorithm, disease types, data type, application, and evaluation metrics. Finally, the paper highlights key results and provides insight into future trends and opportunities in the MLBDD area.

Abbreviation

The paper defines abbreviations for commonly discussed machine-learning and deep-learning methods, models, and related concepts.

  • ANN denotes Artificial Neural Network, while CNN denotes Convolutional Neural Network.
  • DL and DNN denote Deep Learning and Deep Neural Networks, respectively.
  • DT, KMC, KNN, and LR denote Decision Trees, K-Means Clustering, K-Nearest Neighbors, and Logistic Regression.
  • ML, NN, RF, and SVM denote Machine Learning, Neural Network, Random Forest, and Support Vector Machine.

1 Introduction

The introduction frames disease diagnosis as difficult because symptoms can be ambiguous, rare diseases may be overlooked, and healthcare systems may lack sufficient professionals. It presents MLBDD as a time- and cost-efficient approach and defines the review’s scope and questions.

  • 1 Introduction: Ambiguous symptoms and limited health-professional capacity make proper diagnosis difficult, particularly in countries such as Bangladesh and India.
  • 1 Introduction: Rare diseases, unnoticeable symptoms, and mistakenly omitted diseases are identified as diagnostic challenges.
  • 1 Introduction: MLBDD uses healthcare images and tabular patient data to support disease diagnosis.
  • 1 Introduction: The review aims to help researchers and practitioners choose ML and DL methods for rapid and reliable disease detection and classification.
  • 1.1 Motivation: Its questions cover diseases, databases, ML and DL approaches, CNN architectures, and performance evaluation.
  • 1 Introduction: The paper proceeds from ML and DL background and article selection to bibliometric analysis, disease applications, and frequently used ML methods.

2 Basics and Background

The background defines ML as data-driven mathematical and statistical learning that supports forecasting or task-dependent decisions without explicit programming.

  • 2 Basics and Background: Machine learning analyzes data samples with mathematical and statistical approaches so machines can learn without programming.
  • 2 Basics and Background: ML learns from data to forecast outcomes or make decisions according to an assigned task.
  • 2 Basics and Background: ML algorithms are generally classified as supervised, unsupervised, or semi-supervised.

2.1 Machine Learning Algorithms

This section reviews frequently used ML algorithms for disease diagnosis, including tree-based, margin-based, neighborhood, probabilistic, regression, and ensemble approaches.

  • Decision Trees: Decision Trees use divide-and-conquer rules, with leaves representing classes and branches representing feature combinations; regression trees handle continuous variables.
  • Support Vector Machine: Support Vector Machines address classification and regression using a hyperplane to identify clusters.
  • K-Nearest Neighbor: K-Nearest Neighbor assigns class membership through voting and uses Euclidean distance between data samples.
  • Naive Bayes: Naive Bayes estimates membership probabilities for each class and predicts the class with the greatest probability.
  • Logistic Regression: Logistic Regression is a probabilistic classification approach whose Sigmoid function maps real values to the interval from 0 to 1.
  • AdaBoost: AdaBoost combines multiple weak classifiers, weighting difficult samples more heavily and well-classified samples less heavily.

2.2 Deep Learning Overview

Deep Learning uses multiple layers to extract information from inputs and is commonly built on neural-network architectures. CNNs are widely used for image processing and have been applied to disease diagnosis.

  • Deep Learning extracts lower- and higher-level information from image, numerical, and categorical inputs through multiple layers.Each layer transforms its input while learning distinct data attributes.
  • Deep Learning includes supervised, semi-supervised, and unsupervised learning.
  • Artificial Neural Networks underpin most contemporary Deep Learning models, including CNNs, Deep Neural Networks, and Recurrent Neural Networks.Deep Learning may also integrate generative models, Deep Belief Networks, and Boltzmann Machines.
  • 2.2.1 Convolutional Neural Network: CNNs are Artificial Neural Network subclasses extensively used for image processing and biological image detection or recognition.Their applications also include face identification, text analysis, and human organ localization.
  • 2.2.1 Convolutional Neural Network: CNN architectures comprise input, hidden, and output layers and have performed well in disease diagnosis over the last three decades.The passage attributes this development to successive CNN variants proposed since 1989.
  • Figures 2 and 3 provide timelines for well-known CNN models and the development of Machine Learning and Deep Learning algorithms.

2.3 Performance Evaluations

The review describes accuracy, precision, recall, sensitivity, specificity, F-measure, and AUC as performance measures used in disease-diagnosis research. These metrics characterize different aspects of correct, incorrect, positive, and negative predictions.

  • Accuracy measures correctly identified instances among all instances using true-positive, true-negative, false-positive, and false-negative outcomes.The review presents accuracy as a widely used disease-diagnosis performance indicator.
  • Precision measures the proportion of precisely predicted observations among all expected positive observations.
  • Recall measures the proportion of relevant results that the algorithm properly recognizes.
  • Sensitivity measures true positives among total instances, whereas specificity measures correctly identified true negatives.
  • The F1 score is the harmonic mean of precision and recall, with 1 indicating perfect precision and recall.
  • AUC represents model behavior in different situations and is calculated from positive and negative samples ranked by their ratings.The review defines l_p and l_n as positive and negative samples and R_i as the rating of the ith positive sample.

3 Article Selection

The review selected peer-reviewed English journal articles from Scopus and Web of Science using disease-diagnosis and machine-learning-related search terms. After filtering and open-access selection, 1216 full-text articles were retained for bibliographic analysis.

  • The search used Scopus and Web of Science databases and combined disease, diagnosis, and machine-learning-related terms.
  • 16,209 Scopus items and 2,129 Web of Science items were retrieved in the initial search.
  • 9,117 Scopus articles and 1,803 Web of Science articles remained after restricting the period to 2012–2021 and requiring peer-reviewed English papers.
  • 1,216 full-text open-access journal articles were selected for bibliographic analysis, comprising 724 from Scopus and 492 from Web of Science.Article information was imported as CSV data, and duplication functions were used to identify and eliminate duplicates.
  • Exclusion criteria included inaccessible full text, non-human studies, book chapters, reviews, and incomplete test-result information.
  • Figure 4 presents the systematic article-selection flow diagram.

4 Bibliometric Analysis

The bibliometric analysis characterizes MLBDD publications by subject area, keyword networks, publication trends, journals, citations, countries, and authors. It identifies strong concentration in computer science and engineering and highlights leading journals, papers, countries, and contributors.

  • 4.1 Subject Area: Computer science (40%) and engineering (31.2%) were the two dominant subject areas in MLBDD research.
  • 4.2 Co-Occurrence Network: Keyword co-occurrence analysis identified clusters centered on neural networks, decision trees, machine learning, and logistic regression.The neural-network cluster included support vector machine, Parkinson’s disease, and classification.
  • 4.5 Publication by Citations: MLBDD journal publications grew exponentially from 2017, with data covering publications from 2012 to 2021.The paper cautions that the plotted image may not accurately represent MLBDD’s real contribution, although it illustrates influence over time.
  • 4.5 Publication by Citations: IEEE Access and Scientific Reports were the most productive journals, publishing 171 and 133 MLBDD articles, respectively.
  • 4.5 Publication by Citations: Motwani et al. (2017) received the most citations, with 257 citations and 51.4 citations per year, followed by Gray et al. (2013) with 218 citations.Citation counts may differ across Google Scholar and other online databases because indexing procedures and times vary.
  • 4.6 Productive Countries: China contributed the most MLBDD publications with 259 articles, followed by the USA with 139 and India with 103.Four of the ten most productive countries were Asian: China, India, Korea, and Japan.

5 Machine Learning Techniques for Different Disease Diagnosis

This section reviews ML and DL applications across cardiac, kidney, breast, diabetic, Parkinson’s, COVID-19, Alzheimer’s, and other disease diagnoses, highlighting reported performance and limitations.

  • Heart Disease: ML and DL approaches have been applied to cardiac disease diagnosis, including automated systems and deep CNNs for irregular cardiac sounds.A cardiac-sound model reported 0.95 specificity and 0.73 sensitivity in the 2016 PhysioNet competition.
  • Kidney Disease: 98.3% accuracy was reported for SVM kidney-disease classification, while KNN, logistic regression, and decision trees achieved 98.1%, 96.55%, and 94.8%, respectively.Another study found MLP performed best at 98.1% on a comparable CKD dataset.
  • Breast Cancer: ML-based breast-cancer systems include fuzzy-logic computer-aided diagnosis and tuned KNN models for cancer categorization and detection.Grid-search tuning increased KNN accuracy to 94.35% versus approximately 90% with default settings.
  • Diabetes: For diabetes diagnosis, RF outperformed DL and SVM with 83.67% accuracy, while another study reported approximately 98.07% accuracy for DL on the PIMA dataset.The review notes missing discussion of interpretability and performance with imbalanced or incomplete data.
  • Parkinson’s Disease: For Parkinson’s disease, RF achieved 90.26% accuracy and outperformed NB at 69.23%, while a Deep CNN study reported almost 100% accuracy on training and test sets.The Deep CNN study did not discuss overfitting or provide strong interpretation of final predictions.
  • COVID-19: COVID-19 studies used CT and chest X-ray images with ML and DL, reporting 98.5% accuracy for UNet++ and 99% overall accuracy across ten DL models.An Xception model achieved 89.6% accuracy on an imbalanced chest X-ray dataset.
  • Alzheimer’s Disease: Alzheimer’s disease studies reported 85% accuracy with SVM, 83% with decision trees, and 82% for single-subject Alzheimer’s disease and MCI prediction.The review also describes ML and DL applications to eye-movement and skin-disease classification.

6 Algorithm and Dataset Analysis

The review finds that MLBDD research frequently combines multiple algorithms and identifies CNN, SVM, and logistic regression as extensively used individual methods.

  • Algorithm Analysis: Hybrid MLBDD approaches combine multiple algorithms, including Bayesian methods, RF, CNN, SVM, KNN, decision trees, and neural networks.Examples span coronary heart disease, Alzheimer’s disease, heart disease, and anaemia diagnosis.
  • Algorithm Analysis: CNN, SVM, and logistic regression are identified as the most extensively used individual algorithms in the reviewed MLBDD literature.The cited applications include brain-tumor, skin-cancer, liver-disease, and other diagnoses.
  • Algorithm Analysis: The review states that hybrid techniques are more accurate than single ML models.

7 Discussion

The review identifies CNN as the dominant and emerging algorithm in MLBDD, while transfer learning and ensemble methods are also gaining prominence. Its evidence base is constrained by database selection, limited process detail, and the scope of the systematic search.

  • Evidence base: The review draws on bibliometric and dataset-related evidence, including referenced literature and widely used disease-diagnosis datasets.The supplied figure and table captions identify a word cloud of frequently used algorithms and a table of widely used datasets.
  • Algorithms: CNN is the most dominant MLBDD algorithm, with VGG16, VGG19, ResNet50, and UNet++ among the prominent architectures.The review reports CNN performance across image and tabular data and identifies it as more prominent than other ML algorithms.
  • Algorithms: Transfer learning is gaining popularity because it avoids constructing CNN models from scratch and produces better results than typical ML methods.
  • Algorithms: SVM, RF, and DT are commonly used in MLBDD, while several researchers emphasize ensemble techniques.
  • Limitations: The study may omit relevant evidence because it searched only Scopus and WOS, did not examine the ML process in comparable detail, and used a defined systematic-review search scope.The authors note that other databases or broader keywords could produce different findings, while detailed ML processes were outside this review’s scope.

8 Conclusions

The review covers ML/DL-based diagnosis research across several major diseases and concludes that CNN is the most popular method. It also identifies multiclass, imbalanced, missing, explainable, and heterogeneous-data settings as future directions.

  • Scope: The review discusses ML/DL-based diagnosis for heart disease, breast cancer, kidney disease, diabetes, Alzheimer’s disease, and Parkinson’s disease.
  • Conclusion: CNN is identified as the most popular method because of its reported performance in constructing a robust model.
  • Future directions: Future MLBDD research may address multiclass classification with highly imbalanced and missing data.
  • Future directions: The review also points to explainable multiclass classification using XAI and optimization of big data combining numerical, categorical, and image data.
Loading 2112.15538v1…