Machine learning based disease diagnosis: A comprehensive review

In plain words

Doctors and hospitals use computer programs that learn from patient data to help spot diseases early, for example from scans, heart signals or blood test values. This paper is a review of that research. The authors searched two large academic databases for papers from 2012 to 2021, counted who publishes most and where, and then read 102 studies in detail. The studies cover heart, kidney, breast, diabetes, Parkinson's, COVID-19 and Alzheimer's disease, plus a few others. The paper's main takeaway is that convolutional neural networks are the most used method. It also points out gaps: the studies report results inconsistently, rarely explain their models, and seldom check for imbalanced data.

TL;DR

A systematic review of machine learning and deep learning for disease diagnosis, covering 2012 to 2021 §3. It has two parts: a bibliometric analysis of 1,216 papers §4 and a narrative synthesis of 102 full-text papers by disease §5 §6. The review reports no experiments of its own. Use it as a map of the literature and a list of open problems, not as a source of benchmark numbers. The paper names CNNs as the most common method §7; that claim rests on counts of papers, not on controlled comparisons.

Problem

The authors state that diagnosis is hard, that many symptoms are ambiguous, and that misdiagnosis is common §1. They motivate machine learning as a cheaper and faster support for diagnosis §1. Their stated purpose is to help researchers and practitioners choose suitable methods and to find relevant studies §1.1. The review is guided by five questions: which diseases are studied, which databases are used, which ML and DL methods are used, which CNN architectures are common, and whether the reported evaluation is sufficient §1.1.

Method

This is a literature review. It has no algorithm of its own. The card records the review procedure, because a reader who wants to repeat or extend the review needs it.

Selection. The authors search the Scopus and Web of Science (WOS) databases §3.1. The query combines a disease term with diagnosis terms and names of ML and DL methods, applied to titles §3.1. The results are narrowed to 2012 to 2021 and to peer-reviewed English papers §3.2. Only open-access journal articles with full text are kept for the bibliometric analysis §3.3. Two reviewers screen titles and abstracts, and disagreements are settled by discussion §3.3. The full texts of 102 papers are then checked against the inclusion rules, and all 102 are included §3.3. The flow of this process is drawn in a PRISMA-style diagram fig. 4. The steps are summarized below.

# Review selection procedure  [§3]  [fig. 4]
# Input:  title-search records from Scopus and Web of Science
# Output: 102 full-text papers for the narrative synthesis

def select_papers(scopus_records, wos_records):
    # 1. Title search with disease and ML terms  [§3.1]
    candidates = title_search(scopus_records, wos_records)

    # 2. Keep 2012-2021, peer-reviewed, English  [§3.2]
    screened = filter_period_and_language(candidates, 2012, 2021)

    # 3. Keep open-access journal articles with full text  [§3.3]
    full_text_pool = filter_open_access_journals(screened)

    # 4. Remove duplicates in a spreadsheet  [§3.3]
    unique = remove_duplicates(full_text_pool)

    # 5. Two reviewers screen titles and abstracts  [§3.3]
    relevant = screen_title_abstract(unique)

    # 6. Check full texts against the inclusion rules  [§3.3]
    included = check_full_text(relevant)
    return included

Bibliometric analysis. The analysis covers subject area, keyword co-occurrence, publication year, journal, citations, country and author §4. The keyword network was drawn with VOSviewer §4.2. The most-cited list was produced with the R-studio tool §4.5.

Synthesis. The papers are grouped by disease: heart, kidney, breast cancer, diabetes, Parkinson's, COVID-19, Alzheimer's and other diseases §5. For each, the review gives a short narrative and a table of the authors, method, dataset, data type and reported metric tab. 3 to tab. 10. A word cloud counts how often each algorithm appears across the reviewed papers §6 fig. 10.

Algorithms in the background section. The review describes decision trees, SVM, k-nearest neighbors, naive Bayes, logistic regression, AdaBoost and CNNs §2.1 §2.2.1. These descriptions are general and carry no numbers.

Evaluation metrics. The review defines accuracy, precision, recall, sensitivity, specificity, F-measure and AUC §2.3. The equations as printed are below, with the caveats listed under Pitfalls.

ACC = \frac{T_{p}+ T_{N}}{T_{p}+T_{N}+F_{p}+F_{N}} eq. 1

P_n = \frac{T_p}{T_p+F_p} eq. 2

R_c = \frac{T_p}{T_n+F_p} eq. 3 (denominator as printed; see Pitfalls)

S_n = \frac{T_p}{T_p+F_N} eq. 4

S_p = \frac{T_N}{T_N+F_P} eq. 5

F = 2 \times \frac{P_n \times R_c}{P_n + R_c} eq. 6 (text describes it differently; see Pitfalls)

AUC = \frac{\sum R_i(I_p) - I_p\left((I_p+1)/2\right)}{I_p+I_n} eq. 7 (printed with unbalanced parentheses, balanced here for display; not usable as is, see Pitfalls)

Hyperparameters

The review trains no models, so there are no learning hyperparameters. The table lists the parameters of the review itself, which a reproduction would need.

Results

Author-reported. These are counts from the review, not model results. Setup: papers in Scopus and WOS, 2012 to 2021, bibliometric set of 1,216 papers, and narrative set of 102 papers §3.3.

Headline accuracies of the reviewed studies, as the review reports them. The datasets and splits differ, and the review itself warns that these numbers cannot be compared §7.

Paper vs code

No official code. The paper has no code link, no project page and no arXiv comment that points to code. The arXiv metadata has no comments field. A GitHub search for the title and for the authors' names found no repository owned by either author. The three repositories it returned belong to other people and are not authors' code. The review links one GitHub repository, a COVID-19 chest X-ray dataset tab. 11. That is a dataset, not code for this review.

Pitfalls

• Not an experiment. No model is trained or tested. The conclusions are counts of papers and the authors' judgement. The claim that CNNs outperform other ML algorithms §7 describes the reviewed papers, not a controlled comparison.
• Metric equations have errors. Eq. 3 is written with denominator T_n + F_p, while eq. 4 uses T_p + F_N for the same quantity. The text describes recall as the share of relevant results found, which matches eq. 4 more than eq. 3. Do not implement eq. 3 as printed.
• F-measure is described two ways. The text calls the F1 score "the mean of accuracy and recall in a harmonic manner" §2.3. Eq. 6 uses precision and recall. The equation is the one to follow; the text is wrong.
• AUC equation is not usable. Eq. 7 has unbalanced parentheses. It uses I_p and I_n, while the text defines l_p and l_n §2.3. It also uses R_i without defining it beyond "the rating of the i-th positive sample". Take an AUC definition from a source outside this card.
• Sigmoid is called the cost function. The review says logistic regression uses a cost function "often known as the Sigmoid function" §2.1.5. The card records the wording; it does not adopt it.
• Attribution mix-ups between text and tables. The text credits Sun et al. (2021) with GNB, BNB and RF for coronary heart disease, and Bemando et al. (2021) with CNN and SVM for Alzheimer's §6. Tables 3 and 9 credit the reverse: Bemando et al. for the heart-disease methods tab. 3, and Sun et al. for CNN+SVM on Alzheimer's tab. 9. Check the primary paper before quoting an attribution.
• Citation counts disagree. The text gives 218 citations for Gray et al. (2013) §4.5; the table gives 248 tab. 1.
• Counts do not fully reconcile. 1,216 papers go forward, but 1,192 are screened, and the review does not state what removed the other 24 §3.3. The exclusion reasons at full-text stage are listed but not counted §3.3. The PRISMA diagram fig. 4 is the only place that might show the counts.
• The search string has typos. The Boolean query contains "diagnsois" and "Supprot" §3.1. A rerun with the query as printed will not match what the authors matched. The query also includes "adversarial network" and "GAN", which the review does not use later.
• Selection is narrow. Only open-access journal articles were kept for the bibliometric set §3.3. Only Scopus and WOS were searched §7. Results may differ with other sources, as the authors say §7.
• Results are not comparable. Reported accuracies come from private or single datasets, different splits, and mixed metrics (accuracy, AUC, sensitivity, F1). Some table cells are blank or garbled, for example the dataset cell for Magalhaes et al. tab. 10. The review itself says the metrics are inconsistently reported §7.
• Class imbalance is ignored by most studies. The review notes that many reviewed studies use imbalanced data but none discusses it §7. Do not assume the reported accuracies hold on imbalanced clinical data.
• Year trend is weak evidence. The growth in publications from 2017 is shown by a count the authors admit may not reflect real contribution §4.3.
• Background section has errors. Several passages are garbled. For example, the SVM passage says the output "is not non-linearly separable" §2.1.2. The deep-learning categories mix learning paradigms with architectures §2.2. Use the equations and the search procedure, not these passages.
• Figures are software outputs. The keyword network, word cloud and bibliometric maps come from VOSviewer and similar tools §4.2 §6. The paper does not state that the 1,216-record dataset is released §3.

Lineage

• Builds on: none stated. The review cites prior surveys §1 and a heart-disease systematic review by the same two authors, listed in the bibliography as ahsan2021machine (arXiv:2112.06459) §1 §5.1. The paper does not say which earlier work it extends.
• Compared against: not applicable. The review compares no methods of its own.
• Superseded by: not known from content in this wiki.