(S-023) Deep Learning-Based Autoencoder for Outlier Detection in Dosing and Concentration-Time Data: Enabling Reliable Population Pharmacokinetic Modeling

Dongjin Lee – Principal Scientist, Clinical Pharmacology Modeling & Simulation, Amgen; Chih-Wei Lin – Director, Clinical Pharmacology Modeling & Simulation, Amgen

Objectives: Outlier detection in pharmacokinetic (PK) concentration-time data is critical for ensuring the robustness of population pharmacokinetic (PopPK) analysis, as anomalous data points can significantly distort parameter estimation and clinical interpretations. Traditional approaches (e.g., visual inspection, CWRES analysis, DBSCAN, K-Means, One-Class SVM) face limitations due to subjectivity, model dependency, and challenges in handling PK data’s time-series structure and variable sequence lengths. This work aims to address these gaps by proposing a novel automated framework for reliable outlier detection in pharmacometric workflows.

Methods: A transformer-based autoencoder is developed, leveraging self-attention mechanisms to process intensively and sparsely sampled, variable-length PK time-series data following multiple dosing. To overcome limited training data: 1) Synthetic PK datasets are generated using historical dosing regimens and pharmacological principles to ensure biologically plausible concentration-time trajectories. 2) Dose-level subsequence decomposition breaks PK profiles into smaller temporal segments, simplifying patterns and amplifying training data via structured augmentation. The autoencoder learns a compressed representation of normal PK behavior, with outliers flagged via reconstruction error thresholds.

Results: The framework robustly learns the characteristic temporal decay of drug concentrations during dosing cycles. It enables precise identification of doses with anomalous data points exhibiting non-physiological deviations (e.g., upward/downward concentration spikes). By automating outlier detection, the method reduces reliance on subjective manual screening or preliminary model fits, improving scalability and objectivity in PopPK analyses.

Conclusions: This transformer-based autoencoder advances the data quality and analysis robustness by providing an efficient and model-independent outlier detection solution. Future work will explore integration with real-time clinical trial monitoring and multi-modal biomarker data, further enhancing its utility in pharmacometrics.

Citations: [1] Vaswani, A., Shazeer, N., Parmar, N., Uszkoreit, J., Jones, L., Gomez, A. N., ... & Polosukhin, I. (2017). Attention is all you need. Advances in neural information processing systems, 30.

Keywords: outlier detection, deep learning, PK concentration-time data