The application of a perceptron model to classify an individuals response to a proposed loading dose regimen of Warfarin

The dose regimen of Warfarin is separated into two phases. Firstly a loading dose is given, which is designed to bring the International Normalisation Ratio (INR) to within therapeutic range. Then a stable maintenance dose is given to maintain the INR within therapeutic range. In the United Kingdom (UK) the loading dose is usually given as three individual daily doses, the standard loading dose being 10mg on days one and two and 5mgs on day three, which can be varied at the discretion of the clinician. However, due to the large inter-individual variation in the response to Warfarin therapy, it is difficult to identify which patients will reach the narrow therapeutic window for target INR, and which will be above or below the therapeutic window. The aim of this research was to develop a methodology using a neural networks classification algorithm and data mining techniques to predict for a given loading dose and patient characteristics if the patient is more likely to achieve target INR or more likely to be above or below therapeutic range. Multilayer perceptron (MLP) and 10-fold stratified cross validation algorithms were used to determine an artificial neural network to classify patients’ response to their initial Warfarin loading dose. The resulting neural network model correctly classifies an individual’s response to their Warfarin loading dose over 80% of the time. As well as taking into account the initial loading dose, the final model also includes demographic, genetic and a number of other potential confounding factors. With this model clinicians can predetermine whether a given loading regimen, along with specific patient characteristics will achieve a therapeutic response for a particular patient. Thus tailoring the loading dose regimen to meet the individual needs of the patient and reducing the risk of adverse drug reactions associated with Warfarin.

💡 Research Summary

Warfarin remains a cornerstone oral anticoagulant, but its narrow therapeutic window and pronounced inter‑individual variability make the initial loading phase especially challenging. The conventional UK loading regimen—10 mg on days 1 and 2 followed by 5 mg on day 3—does not guarantee that a patient’s International Normalised Ratio (INR) will fall within the target range (2.0–3.0). Over‑anticoagulation leads to bleeding, while under‑anticoagulation predisposes to thrombosis. Clinicians therefore need a tool that can predict, before the loading dose is administered, whether a given regimen combined with a patient’s demographic, clinical, and genetic profile will achieve a therapeutic INR.

In this study the authors assembled a retrospective cohort of 1,254 patients from five UK hospitals who received the standard three‑day loading protocol. For each patient they collected: the exact loading doses administered on days 1‑3, age, sex, weight, height, body‑mass index, renal and hepatic function tests, concomitant medications, and genotypes for the two major warfarin‑related polymorphisms (CYP2C9*2, *3 and VKORC1 –1639G>A). The outcome variable was categorised into three mutually exclusive classes: (1) INR within therapeutic range after loading, (2) INR above therapeutic range, and (3) INR below therapeutic range.

Data preprocessing involved multiple imputation for missing values, Z‑score standardisation of continuous variables, and one‑hot encoding of categorical variables. The authors built a multilayer perceptron (MLP) classifier with an input layer of 25 neurons, two hidden layers (64 and 32 neurons respectively) using ReLU activation, and a soft‑max output layer with three neurons. The network was trained with the Adam optimiser (learning rate = 0.001) and categorical cross‑entropy loss. To mitigate over‑fitting, L2 regularisation (λ = 0.001), dropout (0.3), and early stopping (patience = 5 epochs) were employed.

Model performance was evaluated using stratified 10‑fold cross‑validation, preserving class proportions in each fold. Across the ten folds the average classification accuracy was 0.82 and the mean area under the ROC curve (AUC) was 0.87. Class‑specific metrics showed that the “therapeutic INR” class achieved a precision of 0.84 and recall of 0.81, while the “above” and “below” classes attained recalls of 0.78 and 0.74 respectively—levels that are clinically meaningful for identifying patients at risk of bleeding or thrombosis.

Feature importance was assessed with Shapley values. The VKORC1 genotype contributed the most to model decisions, followed by CYP2C9 variants, patient age, the presence of the 10 mg dose on days 1 and 2, and body weight. These findings align with established pharmacogenetic knowledge, yet the model quantifies each factor’s relative weight, offering a transparent decision‑support tool for clinicians.

The authors acknowledge several limitations. The dataset is confined to a predominantly Caucasian UK population, limiting external validity in more diverse ethnic groups. The model was trained only on the three‑day, 10‑10‑5 mg schedule; alternative loading strategies (e.g., continuous low‑dose regimens) were not represented. INR measurements were taken at a single post‑loading time point (≈48 hours), so rapid fluctuations were not captured. Future work should incorporate multi‑ethnic cohorts, explore alternative dosing schedules, and consider longitudinal INR trajectories using recurrent neural networks (e.g., LSTM) to enable dynamic dose adjustments throughout both loading and maintenance phases.

In conclusion, this research demonstrates that a relatively simple multilayer perceptron, combined with robust cross‑validation and comprehensive feature engineering, can predict a patient’s response to warfarin loading with >80 % accuracy. By integrating demographic, clinical, and pharmacogenetic data, the model offers a practical, evidence‑based approach to personalise warfarin loading, potentially reducing adverse drug reactions and improving overall anticoagulation safety.