Limits...
An Integrative Data Science Pipeline to Identify Novel Drug Interactions that Prolong the QT Interval.

Lorberbaum T, Sampson KJ, Woosley RL, Kass RS, Tatonetti NP - Drug Saf (2016)

Bottom Line: In the EHR at Columbia University Medical Center, we compared the QTc intervals of patients prescribed the flagged drug pairs with patients prescribed either drug individually.We created an AE fingerprint consisting of 13 latently detected side effects.Of 889 pairs flagged in FAERS, eight novel QT-DDIs were significantly associated with prolonged QTc intervals in the EHR and were not due to co-prescribed medications.

View Article: PubMed Central - PubMed

Affiliation: Department of Physiology and Cellular Biophysics, Columbia University, New York, NY, USA.

ABSTRACT

Introduction: Drug-induced prolongation of the QT interval on the electrocardiogram (long QT syndrome, LQTS) can lead to a potentially fatal ventricular arrhythmia known as torsades de pointes (TdP). Over 40 drugs with both cardiac and non-cardiac indications are associated with increased risk of TdP, but drug-drug interactions contributing to LQTS (QT-DDIs) remain poorly characterized. Traditional methods for mining observational healthcare data are poorly equipped to detect QT-DDI signals due to low reporting numbers and lack of direct evidence for LQTS.

Objective: We hypothesized that LQTS could be identified latently using an adverse event (AE) fingerprint of more commonly reported AEs. We aimed to generate an integrated data science pipeline that addresses current limitations by identifying latent signals for QT-DDIs in the US FDA's Adverse Event Reporting System (FAERS) and retrospectively validating these predictions using electrocardiogram data in electronic health records (EHRs).

Methods: We trained a model to identify an AE fingerprint for risk of TdP for single drugs and applied this model to drug pair data to predict novel DDIs. In the EHR at Columbia University Medical Center, we compared the QTc intervals of patients prescribed the flagged drug pairs with patients prescribed either drug individually.

Results: We created an AE fingerprint consisting of 13 latently detected side effects. This model significantly outperformed a direct evidence control model in the detection of established interactions (p = 1.62E-3) and significantly enriched for validated QT-DDIs in the EHR (p = 0.01). Of 889 pairs flagged in FAERS, eight novel QT-DDIs were significantly associated with prolonged QTc intervals in the EHR and were not due to co-prescribed medications.

Conclusions: Latent signal detection in FAERS validated using the EHR presents an automated and data-driven approach for systematically identifying novel QT-DDIs. The high-confidence hypotheses flagged using this method warrant further investigation.

No MeSH data available.


Related in: MedlinePlus

Receiver operating characteristic curves for adverse event fingerprint model and direct evidence control. a Model validation was performed by labeling drug pairs containing a drug with known increased risk of TdP as positive examples. We compared the performance of a model built using latent evidence (AE fingerprint model) to a control model using only direct evidence of QT prolongation. b A second evaluation performed using a list of critical and significant DDIs from the Veteran Affairs Hospital in Arizona. For both validations, the AE fingerprint model significantly outperformed the model built solely with direct evidence. Area under the curve (AUC) is indicated in parentheses. DDIs drug–drug interactions, TdP torsades de pointes, AE adverse event
© Copyright Policy - OpenAccess
Related In: Results  -  Collection


getmorefigures.php?uid=PMC4835515&req=5

Fig2: Receiver operating characteristic curves for adverse event fingerprint model and direct evidence control. a Model validation was performed by labeling drug pairs containing a drug with known increased risk of TdP as positive examples. We compared the performance of a model built using latent evidence (AE fingerprint model) to a control model using only direct evidence of QT prolongation. b A second evaluation performed using a list of critical and significant DDIs from the Veteran Affairs Hospital in Arizona. For both validations, the AE fingerprint model significantly outperformed the model built solely with direct evidence. Area under the curve (AUC) is indicated in parentheses. DDIs drug–drug interactions, TdP torsades de pointes, AE adverse event

Mentions: Of the five fingerprint models evaluated, we found that the model containing 13 features achieved the best performance for drug pair data (area under the curve [AUC] = 0.69 using pairs containing a known CredibleMeds drug) (electronic supplementary Fig. 2); see Table 1 for the list of features that constitute the QT AE fingerprint. Importantly, the QT fingerprint model significantly outperformed the model built using direct evidence, as evaluated by both the CredibleMeds (p = 1.62E−3) and Veteran Affairs (p = 5.22E−10) drug pair standards (Fig. 2). We also compared these models to a previously published additive baseline model for predicting DDIs [19] and found that the latent evidence model outperformed this method (electronic supplementary Fig. 3; CredibleMeds: p < 2.2E−16; Veteran Affairs: p = 2.18E−11). After filtering using both empirical p-values and the 4 % false positive rate cutoff, we obtained 889 putative novel DDIs to be validated in the EHR.Table 1


An Integrative Data Science Pipeline to Identify Novel Drug Interactions that Prolong the QT Interval.

Lorberbaum T, Sampson KJ, Woosley RL, Kass RS, Tatonetti NP - Drug Saf (2016)

Receiver operating characteristic curves for adverse event fingerprint model and direct evidence control. a Model validation was performed by labeling drug pairs containing a drug with known increased risk of TdP as positive examples. We compared the performance of a model built using latent evidence (AE fingerprint model) to a control model using only direct evidence of QT prolongation. b A second evaluation performed using a list of critical and significant DDIs from the Veteran Affairs Hospital in Arizona. For both validations, the AE fingerprint model significantly outperformed the model built solely with direct evidence. Area under the curve (AUC) is indicated in parentheses. DDIs drug–drug interactions, TdP torsades de pointes, AE adverse event
© Copyright Policy - OpenAccess
Related In: Results  -  Collection

Show All Figures
getmorefigures.php?uid=PMC4835515&req=5

Fig2: Receiver operating characteristic curves for adverse event fingerprint model and direct evidence control. a Model validation was performed by labeling drug pairs containing a drug with known increased risk of TdP as positive examples. We compared the performance of a model built using latent evidence (AE fingerprint model) to a control model using only direct evidence of QT prolongation. b A second evaluation performed using a list of critical and significant DDIs from the Veteran Affairs Hospital in Arizona. For both validations, the AE fingerprint model significantly outperformed the model built solely with direct evidence. Area under the curve (AUC) is indicated in parentheses. DDIs drug–drug interactions, TdP torsades de pointes, AE adverse event
Mentions: Of the five fingerprint models evaluated, we found that the model containing 13 features achieved the best performance for drug pair data (area under the curve [AUC] = 0.69 using pairs containing a known CredibleMeds drug) (electronic supplementary Fig. 2); see Table 1 for the list of features that constitute the QT AE fingerprint. Importantly, the QT fingerprint model significantly outperformed the model built using direct evidence, as evaluated by both the CredibleMeds (p = 1.62E−3) and Veteran Affairs (p = 5.22E−10) drug pair standards (Fig. 2). We also compared these models to a previously published additive baseline model for predicting DDIs [19] and found that the latent evidence model outperformed this method (electronic supplementary Fig. 3; CredibleMeds: p < 2.2E−16; Veteran Affairs: p = 2.18E−11). After filtering using both empirical p-values and the 4 % false positive rate cutoff, we obtained 889 putative novel DDIs to be validated in the EHR.Table 1

Bottom Line: In the EHR at Columbia University Medical Center, we compared the QTc intervals of patients prescribed the flagged drug pairs with patients prescribed either drug individually.We created an AE fingerprint consisting of 13 latently detected side effects.Of 889 pairs flagged in FAERS, eight novel QT-DDIs were significantly associated with prolonged QTc intervals in the EHR and were not due to co-prescribed medications.

View Article: PubMed Central - PubMed

Affiliation: Department of Physiology and Cellular Biophysics, Columbia University, New York, NY, USA.

ABSTRACT

Introduction: Drug-induced prolongation of the QT interval on the electrocardiogram (long QT syndrome, LQTS) can lead to a potentially fatal ventricular arrhythmia known as torsades de pointes (TdP). Over 40 drugs with both cardiac and non-cardiac indications are associated with increased risk of TdP, but drug-drug interactions contributing to LQTS (QT-DDIs) remain poorly characterized. Traditional methods for mining observational healthcare data are poorly equipped to detect QT-DDI signals due to low reporting numbers and lack of direct evidence for LQTS.

Objective: We hypothesized that LQTS could be identified latently using an adverse event (AE) fingerprint of more commonly reported AEs. We aimed to generate an integrated data science pipeline that addresses current limitations by identifying latent signals for QT-DDIs in the US FDA's Adverse Event Reporting System (FAERS) and retrospectively validating these predictions using electrocardiogram data in electronic health records (EHRs).

Methods: We trained a model to identify an AE fingerprint for risk of TdP for single drugs and applied this model to drug pair data to predict novel DDIs. In the EHR at Columbia University Medical Center, we compared the QTc intervals of patients prescribed the flagged drug pairs with patients prescribed either drug individually.

Results: We created an AE fingerprint consisting of 13 latently detected side effects. This model significantly outperformed a direct evidence control model in the detection of established interactions (p = 1.62E-3) and significantly enriched for validated QT-DDIs in the EHR (p = 0.01). Of 889 pairs flagged in FAERS, eight novel QT-DDIs were significantly associated with prolonged QTc intervals in the EHR and were not due to co-prescribed medications.

Conclusions: Latent signal detection in FAERS validated using the EHR presents an automated and data-driven approach for systematically identifying novel QT-DDIs. The high-confidence hypotheses flagged using this method warrant further investigation.

No MeSH data available.


Related in: MedlinePlus