Published datasetVerified release

Timing-aware inpatient drug-drug interaction detection and interpretable prioritization for pharmacist review at discharge: supplementary data and code

Interaction rules, detection SQL and aggregate model results from 358,038 MIMIC-IV admissions.

Parsa Lajmiri · Matthew Holman · Thomas Moore · Arran Rounds · Victor Damptey

Published: September 24, 2026
Version: 1.0.0
Files: 6 release files
DOI: 10.55157/csd.ddi.2026.001

About this dataset

Overview

Supplementary data and code for “Timing-Aware Inpatient Drug-Drug Interaction Detection and Interpretable Prioritization for Pharmacist Review at Discharge: A MIMIC-IV Study.” The release contains the exact-match medication mapping, the nine interaction rules, the BigQuery SQL that finds co-active drug pairs in MIMIC-IV v3.1, and the aggregate results behind supplementary tables S1 to S14: model metrics, bootstrap confidence intervals, calibration bins, coefficients, a decision-threshold grid and sensitivity analyses across 358,038 admissions. A data dictionary defines every column. No file contains patient-level data.

Drug interactions Medication safety Clinical decision support Pharmacy practice MIMIC-IV

Inspect before download

Data preview

S14_colisted_comparator.csv · first 3 rows

severity	colisted_pair_admissions	timing_aware_pair_admissions	pct_suppressed	colisted_admissions_alerted	timing_aware_admissions_alerted	pct_fewer_admissions
major	59222	39495	33.3	42900	31169	27.3
moderate	9737	7779	20.1	9588	7699	19.7
minor	15355	9173	40.3	15049	9148	39.2

Read the columns

Data dictionary

data_dictionary.csv

Supplementary_Material.pdf

Supplementary tables S1 to S14 with captions.

Supplementary_Data/S1_mapping_table.csvExact-match medication mapping: 39 source names, 29 generics, 10 classes.4 entries

Name	Type	Units or values	Meaning
src_name	text	lowercase medication name	Medication string as ordered in MIMIC-IV, lowercased. An order maps only on an exact match.
generic	text	29 generic names	Normalized generic drug name.
class	text	antiarrhythmic; antibiotic; anticholinergic; antiplatelet; benzodiazepine; doac; nsaid; opioid; ssri; warfarin	Therapeutic class used by class-level rules.
source	text	seed	Origin of the entry: the predefined seed list.

Supplementary_Data/S2_ruleset.csvThe nine interaction rules as implemented.11 entries

Name	Type	Units or values	Meaning
rule_id	text	R001 to R009	Rule identifier.
rule_type	text	generic+generic; generic+class; class+generic; class+class	Whether each side matches a generic name or a class.
left_generic	text	generic name or blank	Left-side generic, when that side matches a generic.
left_class	text	class or blank	Left-side class, when that side matches a class.
right_generic	text	generic name or blank	Right-side generic, when that side matches a generic.
right_class	text	class or blank	Right-side class, when that side matches a class.
min_age	integer	years; blank = no age limit	Minimum anchor age for the rule to apply. Only R009 is restricted (65).
severity	text	major; moderate; minor	Severity assigned to a match.
evidence	text		Evidence cited: a 2023 AGS Beers Criteria table or clinical consensus.
rationale	text		Plain-language reason the combination is flagged.
source	text		Source of the rule.

Supplementary_Data/S3_detection_logic.sql

Detection and feature-aggregation SQL for MIMIC-IV v3.1 on BigQuery. Load S1 and S2 as tables, replace the project.dataset names and run in order. Requires credentialed MIMIC-IV access; the database itself is not included.

Supplementary_Data/S4_full_model_metrics.csvFull per-model metrics for has_major at every operating point (held-out test set: 71,935 admissions, 6,220 positive).19 entries

Name	Type	Units or values	Meaning
target	text	has_major; readmit_30d (S7a)	Prediction target. has_major = 1 when the admission has at least one major rule match. readmit_30d = 1 when the next admission starts 0 to 30 days after discharge.
feature_set	text	n_drugs_only; core_3; expanded; expanded_no_ddi; expanded_with_ddi; none	Predictors used. n_drugs_only = n_drugs. core_3 = n_drugs, los_hours, age. expanded = core_3 plus n_diagnoses, six comorbidity flags, creatinine_median and albumin_median with missingness indicators. expanded_no_ddi and expanded_with_ddi (readmission only) are expanded without or with the six DDI summaries. none = no-skill reference.
model	text	logistic_regression; random_forest; xgboost; always_negative	Classifier. always_negative predicts no positives and is the no-skill reference.
eval_type	text	cv_mean; cv_std; test	cv_mean and cv_std are the mean and standard deviation across five patient-grouped cross-validation folds of the full cohort (logistic regression only). test is the patient-disjoint 20% held-out set.
threshold_type	text	5fold_cv; default_0.5; train_f1_optimal; train_recall_0.80; predict_all_negative	Operating point. train_f1_optimal maximizes F1 on the training precision-recall curve. train_recall_0.80 is the threshold that reaches training recall of at least 0.80. Both are chosen without the test set.
threshold	number	0 to 1; blank for cross-validation and always-negative rows	Decision threshold on the predicted probability.
n	integer	admissions	Admissions evaluated: the full cohort for cross-validation rows, the held-out set for test rows.
accuracy	number	0 to 1	Share of admissions classified correctly. In cv_std rows this and every metric column hold the standard deviation across folds.
precision	number	0 to 1	True positives divided by predicted positives.
recall	number	0 to 1	True positives divided by actual positives.
specificity	number	0 to 1; blank in cross-validation rows	True negatives divided by actual negatives.
npv	number	0 to 1; blank in cross-validation rows	Negative predictive value: true negatives divided by predicted negatives.
f1	number	0 to 1	Harmonic mean of precision and recall.
roc_auc	number	0 to 1; blank for always-negative test rows	Area under the ROC curve.
avg_precision	number	0 to 1	Average precision (area under the precision-recall curve).
tn	integer	admissions; blank in cross-validation rows	True negatives.
fp	integer	admissions; blank in cross-validation rows	False positives.
fn	integer	admissions; blank in cross-validation rows	False negatives.
tp	integer	admissions; blank in cross-validation rows	True positives.

Supplementary_Data/S5_bootstrap_confidence_intervals.csvBootstrap 95% confidence intervals on the held-out test set.10 entries

Name	Type	Units or values	Meaning
target	text	has_major	Prediction target (see S4).
feature_set	text	n_drugs_only; core_3; expanded; expanded_vs_n_drugs_only	Feature set (see S4). expanded_vs_n_drugs_only rows are paired differences between those two sets.
model	text	logistic_regression; random_forest; xgboost	Classifier.
metric	text	roc_auc; avg_precision; precision@0.5; recall@0.5; f1@0.5; specificity@0.5; delta_roc_auc; delta_avg_precision	Metric estimated. @0.5 metrics use the fixed 0.5 threshold. delta_ metrics are expanded minus n_drugs_only.
point_estimate	number	metric scale	Estimate on the held-out test set.
ci_low	number	metric scale	Lower bound of the 95% bootstrap confidence interval.
ci_high	number	metric scale	Upper bound of the 95% bootstrap confidence interval.
n_bootstrap	integer	1000	Number of bootstrap resamples.
split	text	test	Data split resampled.
notes	text		Resampling method: stratified by label, or paired on the same resampled rows for differences.

Supplementary_Data/S6a_age65_model_metrics.csvS4 repeated for admissions with age 65 or older (145,533 admissions).19 entries

Name	Type	Units or values	Meaning
target	text	has_major; readmit_30d (S7a)	Prediction target. has_major = 1 when the admission has at least one major rule match. readmit_30d = 1 when the next admission starts 0 to 30 days after discharge.
feature_set	text	n_drugs_only; core_3; expanded; expanded_no_ddi; expanded_with_ddi; none	Predictors used. n_drugs_only = n_drugs. core_3 = n_drugs, los_hours, age. expanded = core_3 plus n_diagnoses, six comorbidity flags, creatinine_median and albumin_median with missingness indicators. expanded_no_ddi and expanded_with_ddi (readmission only) are expanded without or with the six DDI summaries. none = no-skill reference.
model	text	logistic_regression; random_forest; xgboost; always_negative	Classifier. always_negative predicts no positives and is the no-skill reference.
eval_type	text	cv_mean; cv_std; test	cv_mean and cv_std are the mean and standard deviation across five patient-grouped cross-validation folds of the full cohort (logistic regression only). test is the patient-disjoint 20% held-out set.
threshold_type	text	5fold_cv; default_0.5; train_f1_optimal; train_recall_0.80; predict_all_negative	Operating point. train_f1_optimal maximizes F1 on the training precision-recall curve. train_recall_0.80 is the threshold that reaches training recall of at least 0.80. Both are chosen without the test set.
threshold	number	0 to 1; blank for cross-validation and always-negative rows	Decision threshold on the predicted probability.
n	integer	admissions	Admissions evaluated: the full cohort for cross-validation rows, the held-out set for test rows.
accuracy	number	0 to 1	Share of admissions classified correctly. In cv_std rows this and every metric column hold the standard deviation across folds.
precision	number	0 to 1	True positives divided by predicted positives.
recall	number	0 to 1	True positives divided by actual positives.
specificity	number	0 to 1; blank in cross-validation rows	True negatives divided by actual negatives.
npv	number	0 to 1; blank in cross-validation rows	Negative predictive value: true negatives divided by predicted negatives.
f1	number	0 to 1	Harmonic mean of precision and recall.
roc_auc	number	0 to 1; blank for always-negative test rows	Area under the ROC curve.
avg_precision	number	0 to 1	Average precision (area under the precision-recall curve).
tn	integer	admissions; blank in cross-validation rows	True negatives.
fp	integer	admissions; blank in cross-validation rows	False positives.
fn	integer	admissions; blank in cross-validation rows	False negatives.
tp	integer	admissions; blank in cross-validation rows	True positives.

Supplementary_Data/S6b_age65_bootstrap_ci.csvS5 repeated for admissions with age 65 or older.10 entries

Name	Type	Units or values	Meaning
target	text	has_major	Prediction target (see S4).
feature_set	text	n_drugs_only; core_3; expanded; expanded_vs_n_drugs_only	Feature set (see S4). expanded_vs_n_drugs_only rows are paired differences between those two sets.
model	text	logistic_regression; random_forest; xgboost	Classifier.
metric	text	roc_auc; avg_precision; precision@0.5; recall@0.5; f1@0.5; specificity@0.5; delta_roc_auc; delta_avg_precision	Metric estimated. @0.5 metrics use the fixed 0.5 threshold. delta_ metrics are expanded minus n_drugs_only.
point_estimate	number	metric scale	Estimate on the held-out test set.
ci_low	number	metric scale	Lower bound of the 95% bootstrap confidence interval.
ci_high	number	metric scale	Upper bound of the 95% bootstrap confidence interval.
n_bootstrap	integer	1000	Number of bootstrap resamples.
split	text	test	Data split resampled.
notes	text		Resampling method: stratified by label, or paired on the same resampled rows for differences.

Supplementary_Data/S7a_readmission_model_metrics.csvExploratory 30-day readmission models; same columns as S4.19 entries

Name	Type	Units or values	Meaning
target	text	has_major; readmit_30d (S7a)	Prediction target. has_major = 1 when the admission has at least one major rule match. readmit_30d = 1 when the next admission starts 0 to 30 days after discharge.
feature_set	text	n_drugs_only; core_3; expanded; expanded_no_ddi; expanded_with_ddi; none	Predictors used. n_drugs_only = n_drugs. core_3 = n_drugs, los_hours, age. expanded = core_3 plus n_diagnoses, six comorbidity flags, creatinine_median and albumin_median with missingness indicators. expanded_no_ddi and expanded_with_ddi (readmission only) are expanded without or with the six DDI summaries. none = no-skill reference.
model	text	logistic_regression; random_forest; xgboost; always_negative	Classifier. always_negative predicts no positives and is the no-skill reference.
eval_type	text	cv_mean; cv_std; test	cv_mean and cv_std are the mean and standard deviation across five patient-grouped cross-validation folds of the full cohort (logistic regression only). test is the patient-disjoint 20% held-out set.
threshold_type	text	5fold_cv; default_0.5; train_f1_optimal; train_recall_0.80; predict_all_negative	Operating point. train_f1_optimal maximizes F1 on the training precision-recall curve. train_recall_0.80 is the threshold that reaches training recall of at least 0.80. Both are chosen without the test set.
threshold	number	0 to 1; blank for cross-validation and always-negative rows	Decision threshold on the predicted probability.
n	integer	admissions	Admissions evaluated: the full cohort for cross-validation rows, the held-out set for test rows.
accuracy	number	0 to 1	Share of admissions classified correctly. In cv_std rows this and every metric column hold the standard deviation across folds.
precision	number	0 to 1	True positives divided by predicted positives.
recall	number	0 to 1	True positives divided by actual positives.
specificity	number	0 to 1; blank in cross-validation rows	True negatives divided by actual negatives.
npv	number	0 to 1; blank in cross-validation rows	Negative predictive value: true negatives divided by predicted negatives.
f1	number	0 to 1	Harmonic mean of precision and recall.
roc_auc	number	0 to 1; blank for always-negative test rows	Area under the ROC curve.
avg_precision	number	0 to 1	Average precision (area under the precision-recall curve).
tn	integer	admissions; blank in cross-validation rows	True negatives.
fp	integer	admissions; blank in cross-validation rows	False positives.
fn	integer	admissions; blank in cross-validation rows	False negatives.
tp	integer	admissions; blank in cross-validation rows	True positives.

Supplementary_Data/S7b_readmission_coefficients.csvLogistic-regression coefficients for readmit_30d with the expanded_with_ddi feature set.3 entries

Name	Type	Units or values	Meaning
feature	text	feature name	Model feature; definitions are listed under Model features.
coefficient	number	log-odds per one native unit	Logistic-regression coefficient on the raw feature scale (no scaling), so magnitudes are not comparable across features.
odds_ratio	number	ratio	exp(coefficient): odds multiplier per one-unit increase.

Supplementary_Data/S8_medication_count_operating_rules.csvRanking the worklist by mapped medication count alone, on the held-out test set.7 entries

Name	Type	Units or values	Meaning
operating_rule	text	n_drugs >= 2 to n_drugs >= 9	Flag admissions with at least this many mapped medications.
alerts	integer	admissions	Admissions flagged.
pct_cohort_flagged	number	percent (0 to 100)	Share of the test set flagged, as a percentage. S10 reports the same measure as a fraction.
true_positives_caught	integer	admissions	Flagged admissions with a major rule match.
recall	number	0 to 1	true_positives_caught divided by all 6,220 positive test admissions.
precision	number	0 to 1	true_positives_caught divided by alerts.
alerts_per_true_positive	number	reviews per catch	Pharmacist reviews needed per captured major rule match.

Supplementary_Data/S9_overlap_sensitivity.csvDetection counts when the co-activity overlap threshold changes.6 entries

Name	Type	Units or values	Meaning
overlap_threshold_hours	number	hours: 0, 1, 2, 4, 8	Minimum overlap of two administration windows for a pair to count as co-active. The main analysis uses 2.
major_pairs	integer	co-active order-line pairs	Major-severity rule matches, counted per pair of overlapping order lines.
moderate_pairs	integer	co-active order-line pairs	Moderate-severity rule matches.
minor_pairs	integer	co-active order-line pairs	Minor-severity rule matches.
admissions_with_major	integer	admissions	Admissions with at least one major match.
pct_change_in_admissions_vs_2h	number	percent, signed	Change in admissions_with_major relative to the 2-hour threshold.

Supplementary_Data/S10_full_threshold_grid.csvExpanded logistic model for has_major across thresholds 0.05 to 0.95, with pharmacist-workload measures.18 entries

Name	Type	Units or values	Meaning
target	text	has_major	Prediction target.
feature_set	text	expanded	Feature set (see S4).
model	text	logistic_regression	Classifier.
threshold_type	text	default_0.5; train_f1_optimal; train_recall_0.80; grid_0.05 to grid_0.95	Named operating point or a grid step.
threshold	number	0 to 1	Decision threshold on the predicted probability.
precision	number	0 to 1	True positives divided by predicted positives.
recall	number	0 to 1	True positives divided by actual positives.
specificity	number	0 to 1	True negatives divided by actual negatives.
f1	number	0 to 1	Harmonic mean of precision and recall.
accuracy	number	0 to 1	Share classified correctly.
tn	integer	admissions	True negatives.
fp	integer	admissions	False positives.
fn	integer	admissions	False negatives.
tp	integer	admissions	True positives.
alerts	integer	admissions	Predicted positives (tp + fp).
true_positives_caught	integer	admissions	Equal to tp.
alerts_per_true_positive	number	reviews per catch	alerts divided by true_positives_caught.
pct_cohort_flagged	number	fraction (0 to 1)	Share of the test set flagged, as a fraction. S8 reports the same measure as a percentage.

Supplementary_Data/S11_calibration_bins.csvReliability curve for the expanded logistic model on the held-out test set.7 entries

Name	Type	Units or values	Meaning
target	text	has_major	Prediction target.
feature_set	text	expanded	Feature set (see S4).
model	text	logistic_regression	Classifier.
bin	integer	0 to 9	Probability bin, ordered from lowest to highest predicted probability.
mean_predicted_prob	number	0 to 1	Mean predicted probability in the bin.
observed_frequency	number	0 to 1	Share of admissions in the bin with a major rule match.
brier_score	number	0 to 1	Brier score of the model on the whole test set, repeated on each row.

Supplementary_Data/S12_full_coefficients.csvLogistic-regression coefficients for has_major with the expanded feature set.3 entries

Name	Type	Units or values	Meaning
feature	text	feature name	Model feature; definitions are listed under Model features.
coefficient	number	log-odds per one native unit	Logistic-regression coefficient on the raw feature scale (no scaling), so magnitudes are not comparable across features.
odds_ratio	number	ratio	exp(coefficient): odds multiplier per one-unit increase.

Supplementary_Data/S13_discharge_active.csvHow many flagged admissions still had a flagged pair co-active in the last 24 hours of the stay.4 entries

Name	Type	Units or values	Meaning
severity	text	major; moderate; minor	Rule severity.
admissions_with_match	integer	admissions	Admissions with at least one match of that severity.
admissions_with_pair_active_within_24h_of_discharge	integer	admissions	Of those, admissions where a flagged pair's co-active window reached the final 24 hours of the stay.
pct	number	percent (0 to 100)	Second count as a percentage of the first.

Supplementary_Data/S14_colisted_comparator.csvWhat the co-activity requirement removes, compared with a checker that flags drugs merely listed together.7 entries

Name	Type	Units or values	Meaning
severity	text	major; moderate; minor	Rule severity.
colisted_pair_admissions	integer	drug pairs per admission	Distinct rule-matched drug pairs per admission when co-listing is enough.
timing_aware_pair_admissions	integer	drug pairs per admission	The same count when the 2-hour co-activity requirement applies.
pct_suppressed	number	percent (0 to 100)	Share of co-listed pairs removed by the requirement.
colisted_admissions_alerted	integer	admissions	Admissions with at least one co-listed match.
timing_aware_admissions_alerted	integer	admissions	Admissions with at least one timing-aware match.
pct_fewer_admissions	number	percent (0 to 100)	Reduction in admissions alerted.

Model featuresAdmission-level predictors named in the feature column of S12 and S7b and in the feature sets.20 entries

Name	Type	Units or values	Meaning
n_drugs	integer	count	Distinct mapped medications in the admission (seed list only, so below the full medication count).
los_hours	number	hours	Span from the first to the last mapped administration; a proxy for medication exposure in the stay.
age	integer	years	MIMIC-IV anchor_age: age in the patient's anchor year.
n_diagnoses	integer	count	Distinct ICD diagnosis codes on the admission.
has_ckd	0/1	flag	Chronic kidney disease: ICD-10 N18 or ICD-9 585.
has_heart_failure	0/1	flag	Heart failure: ICD-10 I50 or ICD-9 428.
has_diabetes	0/1	flag	Diabetes: ICD-10 E10, E11, E13 or ICD-9 250.
has_liver_disease	0/1	flag	Liver disease: ICD-10 K70 to K77 or ICD-9 571.
has_dementia	0/1	flag	Dementia: ICD-10 F01, F02, F03, G30 or ICD-9 290, 3310.
has_copd	0/1	flag	COPD: ICD-10 J44 or ICD-9 491, 492, 496.
creatinine_median	number	mg/dL	Median serum creatinine in the admission (MIMIC-IV itemid 50912); median-imputed when missing.
albumin_median	number	g/dL	Median serum albumin in the admission (MIMIC-IV itemid 50862); median-imputed when missing.
missingindicator_creatinine_median	0/1	flag	1 when no creatinine was measured and the value was imputed.
missingindicator_albumin_median	0/1	flag	1 when no albumin was measured and the value was imputed.
n_major	integer	co-active order-line pairs	Major rule matches. Readmission models only: DDI summaries are outcome-derived for has_major.
n_moderate	integer	co-active order-line pairs	Moderate rule matches (readmission models only).
n_minor	integer	co-active order-line pairs	Minor rule matches (readmission models only).
mean_overlap_hours	number	hours	Mean overlap of the admission's rule-matched pairs (readmission models only).
max_overlap_hours	number	hours	Longest overlap of the admission's rule-matched pairs (readmission models only).
has_any_interaction	0/1	flag	Any rule match of any severity (readmission models only).

Files

timing-aware-ddi-data-1.0.0.zip

57.8 kB

Complete frozen release: 20 source-manifest files plus MANIFEST.json.

SHA-256 e76d5d6d2682ba4d1e5d096c4f7a60e20d1968b1e23d5105670f3febd389c230

Download

data_dictionary.csv

24.6 kB

Every column of every table, with units and codes, and the model-feature definitions.

SHA-256 265652a2e1faf305f04047f365621682f0f68eca1796436bffc18d172cde3509

Download

README.md

3.0 kB

What each file contains, the patient-data note and rights.

SHA-256 4b4925ccc8ce0b7d6619caedc4a3d2d8ba417f2c90564c9c0558197585f999c3

Download

MANIFEST.json

3.5 kB

Byte count and SHA-256 for every file in the source release.

SHA-256 f9f25dc118f6fca90397eb967f220e522d4c7b0f443396b598f95cf2dc3a1339

Download

CITATION.cff

864 bytes

Dataset citation and creator list.

SHA-256 df607e016988963e189f54b8d9cb36b54c4c7bc93283bf38fe9698c53bc1a0ec

Download

deposit-receipt.json

2.1 kB

ConductScience verification and packaging receipt.

SHA-256 b3f080dfcbb09a6df8f1a5e2b4ad638abbccd55475e062c1b07f31bd57e8b9b4

Download

How it was made

Methods

1
Two drugs count as co-active only when their administration windows overlap by at least two hours. Co-active pairs are matched to nine rules derived from the 2023 AGS Beers Criteria (S2) after exact-match medication mapping (S1); S3 holds the BigQuery SQL.
2
Admission-level models (logistic regression, random forest and XGBoost) estimate the chance of a major rule match on a patient-level grouped 80/20 split. S4 to S12 report every metric, interval, threshold and coefficient.
3
The release is aggregate only: no file contains patient-level rows. The source manifest fixes all 20 files by byte count and SHA-256, and the archive adds the manifest itself.

Version history

Version 1.0.0 · current

Published September 24, 2026 from source tag v1.0.0 at commit 88feeb6a3ba8.

Cite this dataset

Lajmiri, P., Holman, M., Moore, T., Rounds, A., & Damptey, V. (2026). Timing-aware inpatient drug-drug interaction detection and interpretable prioritization for pharmacist review at discharge: supplementary data and code (Version 1.0.0) [Data set]. ConductScience Data. https://doi.org/10.55157/csd.ddi.2026.001

Download

https://doi.org/10.55157/csd.ddi.2026.001

DOI registered September 24, 2026 local time (September 24, 2026 at 5:31 PM UTC).

Record metadata

License: No license assigned
Access: Open
Updated: September 24, 2026
Record ID: csdata:timing-aware-ddi:1.0.0

JSON metadata Schema.org Dataset RO-Crate metadata

Rights and reuse

The files are free to read and download. The authors have not assigned a reuse license. Cite the dataset when you use the files. Rerunning the SQL requires your own credentialed MIMIC-IV access through PhysioNet.

Related research

Source database

MIMIC-IV v3.1 (credentialed access)

Rule source

2023 AGS Beers Criteria

Browse all datasets