Skip to main content

Table 1 Summaries of reviewed articles

From: Machine learning-based techniques to improve lung transplantation outcomes and complications: a systematic review

Author

Year

Objective

Population

Data source

Number of inputs

in the final model

ML method

Validation method

Model performance

Result

Troiani and Carlin [29]

2004

Predict the incidence of disease after transplantation

30 subjects

A database of home monitoring data

Six different ordinals

symptom measures

Bayesian models, Markov chain Monte Carlo (MCMC) methods

Cross-validation

For the Bayesian model:

ROC curve < 0.78, Sensitivity = 71.5, Specificity = 91.3

Bayesian models have the best performance in comparison with the Markov model.

Oztekin.A et al. [30]

2009

Predicting the graft survival

16,604 cases

UNOS

283 variables

decision trees, neural networks, logistic regression, Cox regression models

Ten-fold cross-validation, Confusion matrix, Sensitivity, Specificity, Accuracy

The accuracy ranged from 78–86% for logistic regression, from 79–86% for neural networks, and from71–79% for decision trees

The undiscovered relationships were founded among the survival-related variables.

Delen.D et al. [31]

2010

Predict the risk factors for transplantation

310,773 records and 565 variables

UNOS

14 variables

SVM, ANN, MLP, RBF, DT (M5, CART), K-means

MSE, R2, 10-fold cross-validation, sensitivity analysis

SVM with an R2 value of 0.879, Neural network with an R2 value of 0.847

M5 algorithm-based regression tree model with R2 value of 0.785.

Thoracic organ recipients could be classified into ‘‘three’’ risk groups, namely low, medium, and high using a k-means clustering algorithm.

Oztekin.A et al. [32]

2011

Predicting the performance of patients after transplantation

16,771 records and 442 variables.

UNOS

27 variables

Bayesian neural networks

10-fold cross-validation, the R2 value of 0.68

The R2 value of 0.73

The ML models are superior to the existing techniques in terms of both prediction and interpretation capabilities.

Hosseini-Baharanchi. F et al. [19]

2016

Predict Bronchiolitis Obliterans Syndrome incidence

44 LTX recipients who survived ≥ 3 months post-LTX

Masih Daneshvari Hospital database

Five variables

Bayesian non-parametric model

Hazard ratio (HR), Monte Carlo error (MC-error)

MC-error values lower than 0.01

Our analysis revealed that CMV infection was associated with a significant increase in the risk of developing BOS.

Pande.A et al. [20]

2017

Predict pulmonary functions after transplantation

9471 FEV1 evaluations were available from 509 LTx patients

Cleveland Clinic data

17 variables

Gradient Boosting, generalized additive models (GAM), Boosted multivariate trees for spirometry data

cross-validation method

Standardized RMSE (sRMSE) averaged over 100 independent replications.

Developed models illustrate that forced 1-second lung expiratory volume (FEV1) has an important feature-time interaction for lung transplant patients.

Barbosa.E et al. [33]

2017

Predict Bronchiolitis Obliterans Syndrome incidence

176 LTx patients

Cardiothoracic clinic

The predictors were qCT metrics, PFTs, or SQS.

Multivariate logistic regression, SVM

The model’s prediction performance was assessed by AUC or Area under a ROC curve in cross-validated samples

Combination of MVLR and SVM based on PFT values: Max AUC 0.771, whereas models using qCT metrics-only outperformed models: max AUC 0.817

SVM models utilizing PC from qCT outperformed PFT (AUC = 0.817 vs. AUC = 0.767),

Combinations of qCT metrics with PFTs could predict BOS in the LTx group

Oztekin.A et al. [22]

2018

Predict quality of life

60,888 records and 443 features

UNOS

147 input features

Genetic algorithm, GA-kNN, GA-ANN, and GA-SVM models

5-fold cross-validation

Precisionclass1 = 0.992,

Sensitivityclass1 = 0.998, Specificity Class1 = 0.996, F-Measure Class1 = 0.995, And G-Meanclass1 = 0.994, AUC = 85%

Applying GA-ANN, GA-kNN, and GA-SVM models proved that the performance of the lung transplantation process could be improved by the GASVM approach.

Barbosa.E. et al. [21]

2018

Predict Bronchiolitis Obliterans Syndrome incidence

71 LTx patients

Belgium clinic

14 variables

Support vector machines (SVMs)

R2 score

Accuracies for SVM: 83%.

Sensitivity of 73.3% and a specificity of 92.3%

ML utilization showed that qCT metrics predict the eventual onset of BOS.

Mark.E et al. [13]

2019

Predict survival rate

20 000 samples

UNOS

128 variables

Linear regression, Cox proportional hazards model, Random Forest

10-fold cross-validation

RMSE = 5.4, 9.0, and 5.3 for the heart, liver, and lung recipients

For all investigated organs, five-year survival was predicted for the majority of patients.

Braccioni.F et al. [23]

2020

The role of CPET parameters in the development of respiratory symptoms in lung recipients

Twenty-four bilateral LTx recipients.

Tertiary teaching Hospital in Northeast Italy

Nine variables

Forest-Tree as ensemble-of-trees methods

5-fold cross-validation

Bottom boxes, correlation matrix, coefficient score, and box plot for each split.

Muscle pain at peak exercise was associated with basal and exercise- metabolic altered pathways. Dyspnea was associated with the intensity of ventilatory response

Dueñas-Jurado.J et al. [24]

2021

Predict recipient-donor matching

404 lung transplants

Reina Sofı´a University Hospital

48 variables

logistic regression (LR), product unit neural networks (PUNNs)

10-fold cross-validation

Chi-Square, coefficient score, and correlation were investigated to estimate the developed models

The proposed models represent a powerful tool for donor-recipient matching.

Shaish.H et al. [25]

2021

Predict survival rate

221 CT images of ILD patients

Institutional

radiology database

Five variables

CNN, univariable logistic regression model, Cox regression analysis

Five-fold cross-validation

AUC = 0.7417,

Sensitivity = 77%,

Specificity = 66% for the CNN model

Virtual lung wedge resection in patients with ILD can be used as an input to a CNN for predicting the histopathologic UIP pattern and transplant-free survival.

Zafar.F et al. [28]

2022

Predict recipient-donor matching

19,263 eligible double LTxs

UNOS

43 variables

LASSO Cox regression, Random Forest tree, COX regression

Not mentioned

The covariate levels of each recipient and the adjusted total risk score was computed for every recipient and the density plot

LAPT could be effective in matching donor-recipient through lung transplantation.

Su.J et al. [27]

2022

Determine the role of infection in rejection

181 sputum samples from 59 L

Guangzhou Medical University

34 variables

Random Forest models

10-fold cross-validation

AUC for the combination of procalcitonin (PCT), the six-genera, and T-lymphocyte levels were 0.919, 0.898, and 0.895

Airway microbiota along with PCT and T lymphocyte levels were determined as predictive factors in infection and acute rejection.

Stefanuto.P et al. [26]

2022

Predict primary graft dysfunction after lung transplantation

35 lung transplants

recipients

Harefield Hospital

27 variables

Support vector machine (SVM) with a linear kernel, Multivariate analysis of variance (MANOVA)

Not mentioned

For SVM, AUROC = 0.90 and an accuracy of 0.83

Three main chemical classes that were effective in PGD prediction were identified using model development.