Data and Code for SafeDrug replication study

Deep learning replication study on patient medication prediction.
Paper repo

BibTex Citation of source paper

@inproceedings{yang2021safedrug,
    title = {SafeDrug: Dual Molecular Graph Encoders for Safe Drug Recommendations},
    author = {Yang, Chaoqi and Xiao, Cao and Ma, Fenglong and Glass, Lucas and Sun, Jimeng},
    booktitle = {Proceedings of the Thirtieth International Joint Conference on
               Artificial Intelligence, {IJCAI} 2021},
    year = {2021}
}

Table of results

On as-is data processing from paper repo:

Model	DDI	Jaccard	PRAUC	F1-score	Avg # of Meds
LR	0.0776	0.4892	0.648	0.7576	17.1273
RETAIN	0.0813	0.4521	0.6157	0.7549	15.5979
RETAINrev	0.0851	0.4469	0.6109	0.7541	15.2285
GAMENet	0.0588	0.7582	0.8574	0.9348	19.3785
GAMENetPID	0.0149	0.236	0.3738	0.7815	4.9307
SafeDrug_paper	0.062	0.5091	0.6655	0.7626	20.6382
SafeDrug	0.0688	0.525	0.6795	0.7689	21.8901

On data filtered by top-40 side-effects instead of bottom-40, see process_data.ipynb:

Model	DDI	Jaccard	PRAUC	F1-score	Avg # of Meds
LR	0.778	0.4894	0.6483	0.7579	17.146
RETAIN	0.7838	0.4521	0.6157	0.7549	15.5979
RETAINrev	0.7694	0.4473	0.6113	0.7535	15.117
GAMENet	0.1859	0.2387	0.3778	0.7837	4.8137
GAMENetPID	0.7392	0.3813	0.5456	0.6569	10.6871
SafeDrug_paper	0.102	0.1721	0.2896	0.6149	5.517
SafeDrug	0.3542	0.2392	0.3789	0.7088	5.7149

Data download

Obtain permission to download the patient history MIMIC3 csv files:

• prescriptions.csv
• diagnoses_ICD.csv
• procedures_ICD.csv

Obtain drug code mapping files from the paper repo

• ndc2atc_level4.csv https://github.com/sjy1203/GAMENet, RXCUI to ATC (MIMIC)
• drug-atc.csv https://github.com/sjy1203/GAMENet, CID to ATC (side-effects)
• ndc2rxnorm_mapping.csv https://github.com/sjy1203/GAMENet, rxnorm to RXCUI (MIMIC)
• drugbank_drugs_info.csv Paper repo, ATC to SMILES molecule strings
• drug-DDI.csv: Paper repo, side-effect pair tables in CID

Preprocessing code

Run notebook titled process_data.ipynb for data replication attempt
Run notebook titled process_data_paper.ipynb for data as per the paper repo

Training and Evaluation code

Run the eponymous notebooks for each model.
Set the argparser variables as desired.

• args.Test = True for bootstrap evaluation against test dataset
• args.Inf_time = True for inference timing
• both False for training
• args.reverse = 1 to reverse patient visits in RETAIN
• args.beta = 1 to use the SafeDrug PID controller to dynamically adjust DDI_loss contribution
• args.mydata = 1 to use alternate MYDATA_PATH source file paths
• args.smalldata = 1 to use alternate data cut

Mind the model state save and resume paths to proceed to evaluation and inference.

Dependencies

• rdkit-pypi
• memory_profiler
• torch
• pandas
• numpy
• sklearn
• dill
• random
• math
• statistics
• time
• datetime
• sys
• os
• logging
• argparse
• warnings
• collections