Standardization of lithium concentration to the 12-hour level using SimpLi: a simulation study and model validation

Technical Summary: Standardization of Lithium Concentration to the 12-Hour Level Using SimpLi

Problem Statement
Lithium remains a cornerstone treatment for bipolar disorder but possesses a narrow therapeutic window, necessitating strict therapeutic drug monitoring (TDM). The standard clinical protocol requires blood sampling 12 hours after the last evening dose. However, real-world adherence to this 10–14 hour window is poor; studies indicate that approximately half of all lithium measurements occur outside this interval. Deviations from the standard time distort concentration interpretation: late sampling leads to underestimation (risking unnecessary dose increases and toxicity), while early sampling leads to overestimation (risking unjustified dose reductions and therapeutic failure). While machine learning offers a potential solution to standardize these measurements, large, standardized datasets containing precise dosing and sampling times required for model calibration are currently unavailable in Russian clinical practice, and access to detailed international datasets is often restricted.

Methodology
This study employed a simulation-based approach following the ADEMP (Aims, Data-generating mechanisms, Estimands, Methods, Performance measures) framework to develop and validate a multiplicative model named SimpLi.

• Data Generation: A synthetic cross-sectional dataset ($n=1000$) was generated using aggregated descriptive statistics from the Bipolar CHOICE study (Köhler-Forsberg et al., 2025). The simulation targeted specific distributions for:

  • Time since the last dose ($t$): Median of 12 hours (IQR 11–14), with a specific "spike" at exactly 12 hours (27.6% of samples).
  • Serum lithium concentration: Modeled via a square-root transformation to match interval-specific means and standard deviations, then back-transformed.
  • Daily lithium dose: Generated using a Gaussian copula to induce a moderate positive correlation with concentration, then discretized to multiples of 150 mg or 300 mg to reflect clinical tablet strengths.
  • Correlations: The simulation was calibrated to reproduce a target negative correlation between time and concentration ($r \approx -0.30$) and a positive dose-concentration correlation ($r \approx 0.30–0.40$).

• Modeling Strategy: The dataset was split 70/30 (stratified by time intervals) for training and testing. Three modeling steps were evaluated:

  1. Linear Regression (Time only): Predicting concentration based solely on time since the last dose.
  2. Linear Regression (Time + Dose): Adding daily dose as a predictor.
  3. SimpLi (Multiplicative Model with Individual Correction): A two-step procedure where a "Time + Dose" model predicts the expected concentration ($\hat{C}_{meas}$) for a given sampling time. An individual correction factor ($r = C_{meas} / \hat{C}_{meas}$) is calculated to capture the patient's deviation from the population average. This factor is then applied to the model's predicted concentration at 12 hours ($\hat{C}_{12}$) to generate the standardized value: $\hat{C}_{12, standardized} = \hat{C}_{12} \times r$.

• Validation: Performance was assessed using Root Mean Squared Error (RMSE), Mean Absolute Error (MAE), and $R^2$. A proof-of-concept analysis compared SimpLi against the published eLi₁₂ formula using five clinical cases from the source literature.

Key Results

• Simulation Validity: The synthetic dataset successfully reproduced the target distributions, including the exact median and interquartile range of sampling times, interval-specific concentration statistics, and the target time-concentration correlation ($r = -0.29998$).
• Linear Model Limitations: A model using only time as a predictor showed limited accuracy ($R^2 = 0.108$ on the test set). Adding dose improved performance slightly ($R^2 = 0.177$), but both models explained less than 20% of the variance, highlighting significant interindividual heterogeneity.
• SimpLi Performance: The multiplicative model incorporating the individual correction factor demonstrated superior accuracy. In the proof-of-concept analysis of five clinical cases, SimpLi achieved a lower Mean Absolute Error (0.042 mEq/L) compared to the eLi₁₂ formula (0.056 mEq/L). The improvement was most notable in patients with higher lithium concentrations.
• Bias at 12 Hours: Even when sampling occurred exactly at 12 hours, direct prediction without an individual correction factor showed substantial bias (−0.150), supporting the necessity of the individual adjustment factor to account for unmeasured pharmacokinetic variables (e.g., clearance, distribution).

Significance and Claims
The authors claim that SimpLi is a practical tool for standardizing lithium levels when sampling times vary, offering a more personalized approach than universal formulas like eLi₁₂. By separating the "average time trajectory" from the "individual deviation," the model preserves patient-specific pharmacokinetic characteristics, potentially improving accuracy at the extremes of concentration or sampling time.

The study positions SimpLi as a grounded, preliminary solution for clinical decision support. The authors explicitly state that the tool is intended for integration into laboratory information systems or electronic health records to provide clinicians with an automated, corrected 12-hour equivalent value. However, the paper maintains a modest stance, emphasizing that the current validation relies on a single synthetic dataset and a small pilot comparison. The authors assert that external validation on real-world data and robustness testing across diverse clinical scenarios are necessary prerequisites before clinical deployment. The work serves as a proof-of-concept demonstrating the feasibility of statistical standardization in the absence of large, granular real-world datasets.