Prediction of Kidney Transplantation Infection based on Traditional Machine Learning and Deep Learning

Yanrong Li, Shibiao Xu, Li Guo, Qing Zhang, Hailong JinInternational Journal on Bioinformatics & Biosciences 2023 · DOI

Abstract

Infection prediction after kidney transplantation is significant. Most existing models for predicting kidney transplant infection are statistical, unintelligent, and straightforward. The foremost task of this paper is to analyze kidney transplantation data, introduce existing traditional machine learning and deep learning methods from non-temporal and temporal scenarios, respectively, and comprehensively evaluate the predictive power of the methods for kidney transplantation infection. Specifically, in the non-temporal scenario, we use Naïve Bayes, K-Nearest Neighbor, Support Vector Machines, and Random Forest models. In addition, in the temporal scenario, we propose an MTAN model based on a sliding window algorithm to exploit the hidden information of adjacent time series fully. Experimental results show that the kidney transplantation prediction models built by Naïve Bayes and Support Vector Machines have better stability than those constructed by K-Nearest Neighbor and Random Forest. The MTAN model with sliding windows can better mine the hidden temporal information.

Introduction

Kidney transplantation is a critical life-saving procedure, yet the post-operative phase is fraught with infection risks that can jeopardize both the graft and the patient. Traditional monitoring relies on irregular laboratory tests and statistical models like the Cox proportional risk model. However, clinical data presents significant challenges: it is multivariate, irregularly sampled, and highly imbalanced. This study explores modern AI approaches to move beyond traditional statistical methods and improve predictive accuracy for these critical health events.

Methodology

Data Acquisition and Preprocessing

The study utilized a real-world dataset from the Third Medical Center of the PLA General Hospital (2016-2019), encompassing 1,890 patients and 15,841 valid records. The data included 15 physiological variables such as LYM, CD3, and NK cells. A major challenge was the extreme class imbalance, with only 2.95% of records labeled as "infected."

To address this, the researchers employed:

  • Class Imbalance Correction: Techniques including SMOTE (Synthetic Minority Over-sampling Technique), Undersampling, and EasyEnsemble were used to balance the training sets.
  • Feature Selection: REFCV and L1 regularization narrowed the 15 variables down to the 6 most predictive ones.

Non-Temporal Scenario

In this scenario, lab results were treated as independent data points. Four classic machine learning models were evaluated: Naïve Bayes (NB), K-Nearest Neighbor (KNN), Support Vector Machines (SVM), and Random Forest (RF).

Traditional Machine Learning Pipeline

Temporal Scenario

To capture the longitudinal nature of patient health, the authors utilized Multi-Time Attention Networks (MTAN). Unlike standard neural networks, MTAN handles irregular time series using learnable time embeddings. A sliding window algorithm was integrated to segment patient histories into manageable chunks for the MTAN encoder, allowing the model to exploit hidden patterns in adjacent time series.

MTAN Architecture for Irregularly Sampled Time Series

Results and Discussion

The Accuracy Trap and Model Stability

Initial tests on imbalanced data yielded over 97% accuracy but 0% recall, highlighting the necessity of balancing techniques. In the non-temporal analysis, simpler models like Naïve Bayes and SVM demonstrated greater stability and effectiveness in classifying positive samples compared to Random Forest, which was prone to overfitting on noisy clinical data. SVM, combined with the EasyEnsemble method, achieved the best performance with an AUROC of 71% and a recall of 60.61%.

Temporal Modeling Performance

The MTAN model's ability to treat time as a continuous variable proved advantageous for hospital data. The inclusion of a sliding window significantly boosted sensitivity. Specifically, moving from a standard imbalanced approach to an Undersampling + Sliding Window configuration increased the recall from 0% to 65%. While the MTAN model excelled at mining temporal information, it faced a trade-off with a lower F1 score due to increased false positives—a common occurrence when prioritizing high recall in medical screenings.

Conclusion

This research demonstrates that combining traditional machine learning with deep learning temporal models significantly improves infection prediction in kidney transplant patients. While the MTAN model with sliding windows effectively mines temporal data, future work is needed to address "internal imbalance" within individual patient time series and to explore model ensembling to combine the stability of traditional ML with the depth of deep learning.

Summary updated

Abstract

nfection prediction after kidney transplantation is significant. Most existing models for predicting kidney transplant infection are statistical, unintelligent, and straightforward. The foremost task of this paper is to analyze kidney transplantation data, introduce existing traditional machine learning and deep learning methods from non-temporal and temporal scenarios, respectively, and comprehensively evaluate the predictive power of the methods for kidney transplantation infection. Specifically, in the non-temporal scenario, we use Naïve Bayes, K-Nearest Neighbor, Support Vector Machines, and Random Forest models. In addition, in the temporal scenario, we propose an MTAN model based on a sliding window algorithm to exploit the hidden information of adjacent time series fully. Experimental results show that the kidney transplantation prediction models built by Naïve Bayes and Support Vector Machines have better stability than those constructed by K-Nearest Neighbor and Random Forest. The MTAN model with sliding windows can better mine the hidden temporal information.
BibTeX
@article{li2023prediction,
  author = {Yanrong Li and Shibiao Xu and Li Guo and Qing Zhang and Hailong Jin},
  title = {{Prediction of Kidney Transplantation Infection based on Traditional Machine Learning and Deep Learning}},
  journal = {International Journal on Bioinformatics \& Biosciences},
  year = {2023},
  doi = {10.5121/ijbb.2023.13102},
}
Figure