Machine Learning-Based Prediction Method for Tremors Induced by Tacrolimus in the Treatment of Nephrotic Syndrome

Bing Shao, Youyang Qu, Wei Zhang, Haihe Zhan, Zerong Li, Xingyu Han, Mengchao Ma, Zhimin DuFrontiers in Pharmacology 2022 · DOI

TL;DR

Researchers have developed a neural network-based model to predict which patients with nephrotic syndrome (NS) are likely to develop tremors when treated with the immunosuppressant tacrolimus. By narrowing down 64 clinical variables to just seven key biomarkers, the model achieves high sensitivity, potentially allowing clinicians to personalize drug dosages and improve patient quality of life.

What Problem Does This Paper Address?

Tacrolimus is a cornerstone of immunosuppressive therapy, particularly for steroid-resistant nephrotic syndrome. However, its narrow therapeutic index means side effects are common. Tremors are the most frequent neurotoxic manifestation, affecting up to 56% of patients. These tremors are not just a nuisance; they significantly hamper the daily activities of young, active patients.

While high doses are known to cause neurotoxicity, tremors often persist even at low doses in NS patients. Currently, clinicians lack a reliable way to predict which patients are at risk before treatment begins. This paper addresses that gap by using machine learning to identify the specific clinical "signatures" that precede tacrolimus-induced tremors.

Methods

The authors employed a two-stage machine learning pipeline:

  1. Feature Selection (Recursive Feature Elimination - RFE): Starting with 64 variables (demographics and laboratory values), the team used RFE to prune the list. This greedy algorithm iteratively removes the least important features based on their impact on model accuracy until an optimal subset is found.
  2. Classification (Neural Network - NN): A single-hidden-layer neural network was trained on the selected features. The architecture used a ReLU activation function for the hidden layer (tuned to 8 neurons) and a sigmoid function for the output layer to produce a probability score between 0 and 1.

The model was trained using 10-fold cross-validation to ensure robustness and was compared against a Random Forest (RF) model to benchmark performance.

Datasets

The study utilized a retrospective dataset from the Second Affiliated Hospital of Harbin Medical University:

  • Total Cohort: 252 patients diagnosed with NS between 2013 and 2017.
  • Incidence: 39 patients (15.5%) experienced tremors.
  • Split: 181 patients were assigned to the training set (including 32 tremor cases), and 71 patients were reserved for an external validation set (including 7 tremor cases).
  • Features: 64 initial variables including blood counts, liver/kidney function tests, and coagulation markers.

Key Insights and Ideas

  • The "Magic Seven": The RFE process identified seven indispensable variables for prediction: creatinine (CREA), D-dimer, total protein (TP), calcium ion (Ca), platelet distribution width (PDW), serum potassium (Ka), and fibrinogen (FIB).
  • Physiological Synergy: The authors suggest these biomarkers aren't random; they reflect the unique physiological state of NS. For instance, because tacrolimus is highly protein-bound, the low protein levels (hypoproteinemia) and altered coagulation states (high FIB and D-dimer) typical of NS likely change the drug's pharmacokinetics, increasing neurotoxicity risk.
  • Neural Networks vs. Random Forest: Interestingly, the Random Forest model failed significantly on this dataset, predicting zero tremor cases in the confusion matrix. The Neural Network proved much better at handling the imbalanced nature of the data (where non-tremor cases far outnumber tremor cases).

Key Results

The neural network model demonstrated excellent predictive power:

  • Training Set:
    • AUC: 0.9726 (indicating near-perfect separation between classes).
  • Accuracy: 93.45%
    • Sensitivity: 97.12% (excellent at catching true tremor cases).
  • External Validation Set:
    • Accuracy: 82.14%
    • Sensitivity: 83.78%
    • Specificity: 70.00%

These results suggest that the model generalizes well to new patients, maintaining high sensitivity even in a real-world validation scenario.

Limitations and Future Work

The authors acknowledge several hurdles that need to be cleared before clinical adoption:

  • Sample Size: The number of tremor cases (39) is relatively small, which can lead to overfitting.
  • Missing Data: Important variables like blood lipid levels and specific drug-drug interactions were not included due to lack of data in the electronic records.
  • Single-Center Study: The data comes from one hospital in China; multi-center studies are needed to ensure the model works across different populations and clinical practices.
  • Black Box Nature: While the model is accurate, the exact biological mechanism of how these seven variables interact to cause tremors remains a hypothesis that requires further experimental validation.

Summary updated

Abstract

Tremors have been reported even with a low dose of tacrolimus in patients with nephrotic syndrome and are responsible for hampering the day-to-day work of young active patients with nephrotic syndrome. This study proposes a neural network model based on seven variables to predict the development of tremors following tacrolimus. The sensitivity and specificity of this algorithm are high. A total of 252 patients were included in this study, out of which 39 (15.5%) experienced tremors, 181 patients (including 32 patients who experienced tremors) were randomly assigned to a training dataset, and the remaining were assigned to an external validation set. We used a recursive feature elimination algorithm to train the training dataset, in turn, through 10-fold cross-validation. The classification performance of the classifer was then used as the evaluation criterion for these subsets to find the subset of optimal features. A neural network was used as a classification algorithm to accurately predict tremors using the subset of optimal features. This model was subsequently tested in the validation dataset. The subset of optimal features contained seven variables (creatinine, D-dimer, total protein, calcium ion, platelet distribution width, serum kalium, and fibrinogen), and the highest accuracy obtained was 0.8288. The neural network model based on these seven variables obtained an area under the curve (AUC) value of 0.9726, an accuracy of 0.9345, a sensitivity of 0.9712, and a specificity of 0.7586 in the training set. Meanwhile, the external validation achieved an accuracy of 0.8214, a sensitivity of 0.8378, and a specificity of 0.7000 in the validation dataset. This model was capable of predicting tremors caused by tacrolimus with an excellent degree of accuracy, which can be beneficial in the treatment of nephrotic syndrome patients.
BibTeX
@inproceedings{shao2022machine,
  author = {Bing Shao and Youyang Qu and Wei Zhang and Haihe Zhan and Zerong Li and Xingyu Han and Mengchao Ma and Zhimin Du},
  title = {{Machine Learning-Based Prediction Method for Tremors Induced by Tacrolimus in the Treatment of Nephrotic Syndrome}},
  booktitle = {Frontiers in Pharmacology},
  year = {2022},
  doi = {10.3389/fphar.2022.708610},
}
Figure