TL;DR
This paper introduces a machine learning approach to predict the need for a reduced drug dose at the initial prescription, using only prior Electronic Health Record (EHR) data. By analyzing patient phenotypes, the system can accurately predict dose reductions for a significant number of drugs, potentially preventing adverse reactions and improving personalized medicine.
What Problem Does This Paper Address?
The core challenge addressed is achieving "precision medicine" – prescribing the right drug at the right dose for each individual. A major hurdle is the significant inter-individual variability in how patients respond to medications. This variability can lead to adverse drug reactions (ADRs) or suboptimal treatment efficacy. While genetics plays a role, it's not always readily available. This paper tackles the problem of predicting when a reduced drug dose might be necessary before the first prescription, using readily available patient history from EHRs, specifically focusing on drugs metabolized by the P450 enzyme family, which are known for their variable responses.
Methods
The researchers developed a machine learning pipeline to predict dose reductions. The process involved:
- Data Preparation: Extracting diagnostic codes, clinical notes (conditions), and laboratory orders from EHRs.
- Outcome Identification: Defining "dose change intervals" – instances where a patient was prescribed the same drug ingredient twice within 20 days, with either a reduced dose (dose reduction), an increased dose (dose increase), or the same dose (dose continuation).
- Phenotype Profile Construction: For each drug, identifying features (diagnostic codes, conditions, lab orders) that were significantly enriched in patients who needed a dose reduction compared to those who had dose continuation. This involved statistical filtering (hypergeometric test, Bonferroni correction) and ontology expansion to capture broader relationships.
- Model Training: Using Random Forest classifiers to predict dose reduction based on these phenotype profiles.
- Evaluation: Employing two validation strategies: 10-fold cross-validation and a "hold last year out" approach to simulate real-world future prediction.
Datasets
The study utilized de-identified EHR data from the STRIDE clinical data warehouse at Stanford University, covering patients from 2008 to 2014. This dataset comprised over 1.25 million patients, 49 million visits, 27 million clinical notes, and 19 million drug orders, including approximately 2.9 million orders for P450-drugs. No new benchmark datasets were introduced; the focus was on leveraging existing EHR data.
Key Insights and Ideas
- Phenotypic History as a Predictor: The study demonstrates that a patient's historical phenotypic data (diagnoses, conditions, lab results) prior to a prescription can be a powerful predictor of their response to that drug, specifically the need for dose reduction.
- Dose Reduction as a Surrogate for Adverse Reactions: For many drugs, dose reduction is only recommended in cases of adverse reactions. The ability to predict dose reduction can therefore serve as an indirect way to predict and potentially prevent ADRs.
- Lab Orders as Key Predictors: Laboratory test orders emerged as the most important feature type for predicting dose reductions, contributing more significantly than diagnostic codes or condition mentions.
- Focus on Dose Reduction, Not Increase: The model was highly effective at predicting dose reductions but struggled to predict dose increases, suggesting that the EHR data contains more observable indicators for negative responses than for lack of positive response.
Key Results
- The model successfully predicted dose reductions for 23 out of 34 drugs and 22 out of 23 drug classes with an AUC-ROC generally between 0.70 and 0.95.
- In the "hold last year out" evaluation (a more stringent test), the model achieved an average AUC-ROC of 0.68 and F-measure of 0.64.
- For specific drugs like Tacrolimus, the model achieved high performance with AUC-ROC of 0.94 and F-measure of 0.89 in 10-fold cross-validation.
- The study identified specific phenotypic profiles associated with dose reductions for drugs like Tacrolimus (linked to urine drug analysis), Labetalol (linked to renal failure diagnostics), and Warfarin (linked to candidiasis and potential drug interactions).
Limitations and Future Work
The authors acknowledge several limitations:
- Dosing Changes as Part of Protocol: For some drugs (e.g., glucocorticoids), dose reduction is standard practice and not necessarily indicative of an adverse reaction. The model's predictions in these cases might not directly inform patient care beyond standard guidelines.
- Ignoring Discontinuation/Switching: The model doesn't account for instances where adverse reactions are managed by discontinuing or switching medications, which could impact prediction accuracy.
- Co-administration of Drugs: The study did not consider the impact of multiple drug interactions.
- Inability to Predict Dose Increases: The model's failure to predict dose increases is a significant limitation, possibly due to a lack of observable features in EHRs for "absence of drug response."
- Limited Feature Set: The model did not utilize actual lab test values, only orders, which could provide richer information.
- Temporal Information Loss: The encoding of data might lose crucial temporal ordering of events.
Future work could involve:
- Next-Generation Algorithms: Employing recurrent neural networks (RNNs) for better handling of temporal data.
- Additional Confounding Factors: Incorporating more complex factors and potentially genetic data.
- Combining Approaches: Integrating data-driven methods with mechanistic understanding and physiological knowledge.
- Utilizing Lab Values: Incorporating actual laboratory test results to improve prediction, especially for dose increases.
- Prospective Validation: Further validating the model in real-world clinical settings.
Summary updated