PK-RNN-V E: A deep learning model approach to vancomycin therapeutic drug monitoring using electronic health record data

Masayuki Nigo, Hong Thoai Nga Tran, Ziqian Xie, Han Feng, Bingyu Mao, Laila Rasmy, Hongyu Miao, Degui Zhi

Abstract

Vancomycin is a commonly used antimicrobial in hospitals, and therapeutic drug monitoring (TDM) is required to optimize its efficacy and avoid toxicities. Bayesian models are currently recommended to predict the antibiotic levels. These models, however, although using carefully designed lab observations, were often developed in limited patient populations. The increasing availability of electronic health record (EHR) data offers an opportunity to develop TDM models for real-world patient populations. Here, we present a deep learning-based pharmacokinetic prediction model for vancomycin (PK-RNN-V E) using a large EHR dataset of 5,483 patients with 55,336 vancomycin administrations. PK-RNN-V E takes the patient’s real-time sparse and irregular observations and offers dynamic predictions. Our results show that RNN-PK-V E offers a root mean squared error (RMSE) of 5.39 and outperforms the traditional Bayesian model (VTDM model) with an RMSE of 6.29. We believe that PK-RNN-V E can provide a pharmacokinetic model for vancomycin and other antimicrobials that require TDM.

Introduction: The Clinical Need for Precision Dosing

Vancomycin is a critical antibiotic with a narrow therapeutic window. Sub-therapeutic levels risk treatment failure and bacterial resistance, while supra-therapeutic levels can lead to nephrotoxicity. To navigate this, clinicians rely on Therapeutic Drug Monitoring (TDM) to adjust dosages based on measured serum concentrations.

Figure 1-A: Clinical scenario of vancomycin dosing and monitoring, illustrating the relationship between dose adjustments, serum levels, and kidney function markers like creatinine.

Current "gold standard" Bayesian models (VTDM) often fall short in real-world clinical settings. These models are typically built on small, homogenous cohorts and rely on a limited set of variables such as age, weight, and serum creatinine. Consequently, they struggle to provide accurate predictions for "unstable" patients—such as those in Intensive Care Units (ICUs)—whose hemodynamic status and renal function fluctuate rapidly.

Methods: Integrating Deep Learning with Pharmacokinetics

To address these limitations, we developed PK-RNN-V E, a hybrid model that combines the sequential processing power of Recurrent Neural Networks (RNNs) with established pharmacokinetic (PK) principles.

Model Architecture

The core of PK-RNN-V E is a Gated Recurrent Unit (GRU) that processes high-dimensional EHR data, including 40 continuous variables (labs and vitals) and 8-dimensional embeddings for over 300 medications. Unlike standard "black box" AI, the RNN does not predict drug levels directly. Instead, it estimates two latent biological parameters: the elimination rate ($k$) and the volume of distribution ($V$).

Figure 1-B: The PK-RNN-V E architecture, showing how EHR data is embedded and processed through an RNN to predict PK parameters (k, v) which are then used in a standard PK formula.

These parameters are then fed into a one-compartment pharmacokinetic differential equation to calculate the final drug concentration. This "PK head" ensures that all predictions remain biologically plausible. The "E" in the model name refers to an ensemble approach (averaging 10 models), and the system incorporates a feedback mechanism where the first measured vancomycin level is used to refine the model’s internal state for a specific patient.

Data and Results

The model was trained and validated using a retrospective dataset from the Memorial Hermann Health System, encompassing 5,483 adult patients and over 55,000 vancomycin administrations.

PK-RNN-V E demonstrated superior predictive accuracy compared to the traditional Bayesian baseline:

  • Overall Accuracy: The model achieved a Root Mean Squared Error (RMSE) of 5.39, significantly lower than the 6.29 achieved by the VTDM model.
  • Robustness: The model maintained high performance across challenging subgroups, including elderly patients and those with high BMI, where traditional equations often fail.
  • Feature Inference: Ablation studies revealed that the model could accurately predict drug levels even when primary kidney function markers (creatinine/GFR) were missing, likely by inferring renal health from other physiological markers like blood pressure and heart rate.

Discussion and Conclusion

PK-RNN-V E represents a significant step forward in precision dosing. By leveraging the breadth of EHR data and the structural constraints of pharmacology, the model provides a more flexible and accurate tool for TDM than traditional Bayesian methods.

While the results are promising, future work is needed to validate the model across different healthcare systems and in highly complex populations, such as patients on dialysis. Ultimately, prospective clinical trials will be necessary to determine if the implementation of PK-RNN-V E leads to improved patient outcomes, such as reduced rates of acute kidney injury and more effective infection control.

Summary updated

BibTeX
@article{nigo2022pk,
  author = {Masayuki Nigo and Hong Thoai Nga Tran and Ziqian Xie and Han Feng and Bingyu Mao and Laila Rasmy and Hongyu Miao and Degui Zhi},
  title = {{PK-RNN-V E: A deep learning model approach to vancomycin therapeutic drug monitoring using electronic health record data}},
  journal = {J. Biomed. Informatics},
  year = {2022},
  doi = {10.1016/j.jbi.2022.104166},
}
Figure