Skip to main page content
U.S. flag

An official website of the United States government

Dot gov

The .gov means it’s official.
Federal government websites often end in .gov or .mil. Before sharing sensitive information, make sure you’re on a federal government site.

Https

The site is secure.
The https:// ensures that you are connecting to the official website and that any information you provide is encrypted and transmitted securely.

Access keys NCBI Homepage MyNCBI Homepage Main Content Main Navigation
. 2021 Mar 5;18(5):2600.
doi: 10.3390/ijerph18052600.

Multi-Drug Featurization and Deep Learning Improve Patient-Specific Predictions of Adverse Events

Ioannis N Anastopoulos  1   2 , Chloe K Herczeg  2 , Kasey N Davis  2 , Atray C Dixit  2
Affiliations

Multi-Drug Featurization and Deep Learning Improve Patient-Specific Predictions of Adverse Events

Ioannis N Anastopoulos et al. Int J Environ Res Public Health. .

Erratum in

Abstract

While the clinical approval process is able to filter out medications whose utility does not offset their adverse drug reaction profile in humans, it is not well suited to characterizing lower frequency issues and idiosyncratic multi-drug interactions that can happen in real world diverse patient populations. With a growing abundance of real-world evidence databases containing hundreds of thousands of patient records, it is now feasible to build machine learning models that incorporate individual patient information to provide personalized adverse event predictions. In this study, we build models that integrate patient specific demographic, clinical, and genetic features (when available) with drug structure to predict adverse drug reactions. We develop an extensible graph convolutional approach to be able to integrate molecular effects from the variable number of medications a typical patient may be taking. Our model outperforms standard machine learning methods at the tasks of predicting hospitalization and death in the UK Biobank dataset yielding an R2 of 0.37 and an AUC of 0.90, respectively. We believe our model has potential for evaluating new therapeutic compounds for individualized toxicities in real world diverse populations. It can also be used to prioritize medications when there are multiple options being considered for treatment.

Keywords: FDA FAERS; UK Biobank; adverse events; graph convolution; neural networks; real world evidence.

PubMed Disclaimer

Conflict of interest statement

A.D. and K.D. are equity holders in Coral Genomics, Inc.

Figures

Figure 1
Figure 1
Overview of our approach: (A) Integration of multiple real world evidence databases including demographic, medication, and genetic information; (B) A machine learning model to predict adverse events is constructed.
Figure 2
Figure 2
An overview of the multi-drug GCN architecture: (A) A standard GCN applied to a chemical structure creates bond and atom-level features, and an atom-level connectivity matrix to describe the molecule. Graph convolutions are performed to learn new feature representations that learn local structures that can be used to predict chemical properties (B) Our multi-drug GCN architecture concatenates the bond and atom- level features and creates a block diagonal connectivity matrix that represents the set of molecules an individual is taking. In a generalization of the single molecule GCN, the multi-drug GCN aggregates information from local structures across all molecules to predict multi-drug properties. We highlight the featurization of an example patient currently taking simvastatin (red pill), ibuprofen (green pill), and metformin (blue pill).
Figure 3
Figure 3
Predictive utility of various features and model architectures for predicting adverse events in the UKBB dataset: (A) Results for hospitalization (top) and death (bottom). Red bar corresponds to neural network architecture, blue bar corresponds to linear model. Y-axis is an R2 measure of model performance on predicting log10 (hospitalization + 1) (top) and AUC for predicting death (bottom). X-axis contains various combinations of features used in the model. Error bars correspond to 95% confidence interval derived from bootstrapping on 5-fold cross-validation (each fold contains 58,312 patients). (B) Feature weights (attributions) for each the top 10 most important non-drug structure features in the integrated model for predicting hospitalization in the UKBB dataset (top) and death (bottom). Error bars correspond to ±1 s.d. (C) Bar plot comparing results of using single drug features to using multi-drug features alone for predicting hospitalization. Error bars correspond to 95% confidence interval derived from bootstrapping on 5-fold cross-validation (each fold contains 42,114 patients).
Figure 4
Figure 4
Performance comparisons on the FAERS dataset. (A) Predictive utility of various features and model architectures for predicting adverse events in the FAERS dataset. X-axis labels correspond to adverse event categories for a particular case. Y-axis is the AUC at predicting each of the labels. Colors correspond to various feature subsets tested. Error bars correspond to 95% confidence interval derived from bootstrapping on 5-fold cross-validation (each fold contains 28,682 records). (B) Power analysis demonstrating improvement in performance as a function of the number of patient records examined. Blue corresponds to hospitalization model performance and orange corresponds to performance of model predicting death. X-axis is log10 (number of records) Y-axis is AUC. Shaded error region corresponds to 95% confidence interval derived from bootstrapping on 5-fold cross-validation in a subsampled dataset corresponding to the X-axis location. (C) Plot demonstrating relationship between model error across all outcomes and age, (D) average molecular weight of drugs patient is taking, and (E) patient sex.
See this image and copyright information in PMC

Similar articles

See all similar articles

Cited by

References

    1. Clark L.T., Watkins L., Piña I.L., Elmer M., Akinboboye O., Millicent G., Jamerson B., McCullough C., Pierre C., Polis A.B., et al. Increasing Diversity in Clinical Trials: Overcoming Critical Barriers. Curr. Probl. Cardiol. 2019;44:148–172. doi: 10.1016/j.cpcardiol.2018.11.002. - DOI - PubMed
    1. Mak W.W.S., Law R.W., Alvidrez J., Pérez-Stable E.J. Gender and Ethnic Diversity in NIMH-Funded Clinical Trials: Review of a Decade of Published Research. Adm. Policy Ment. Heal. Ment. Heal. Serv. Res. 2007;34:497–503. doi: 10.1007/s10488-007-0133-z. - DOI - PubMed
    1. Ramamoorthy A., Pacanowski M., Bull J., Zhang L. Racial/Ethnic Differences in Drug Disposition and Response: Review of Recently Approved Drugs. Clin. Pharmacol. Ther. 2015;97:263–273. doi: 10.1002/cpt.61. - DOI - PubMed
    1. Ksenia J.G., Carvalhoc N.R., Chipmand K.J., Denslowe N.D., Halder M., Murphy C.A., Roelofs D., Rolaki A., Schirmer K., Watanabek K.H. Development and Application of the Adverse Outcome Pathway Framework for Understanding and Predicting Chronic Toxicity: I. Challenges and Research Needs in Ecotoxicology. Chemosphere. 2015;120:764–777. - PubMed
    1. Ngufor C., Wojtusiak J., Pathak J. A Systematic Prediction of Adverse Drug Reactions Using Pre-Clinical Drug Characteristics and Spontaneous Reports. Int. Conf. Healthc. Inform. 2015 doi: 10.1109/ICHI.2015.16. - DOI

Publication types

Substances

LinkOut - more resources