README.md

---
license: cc-by-4.0
task_categories:
  - text-classification
  - question-answering
  - text-retrieval

language:
  - en

tags:
  - medical

pretty_name: MedLens

size_categories:
  - 100K<n<1M
---

# Dataset Card for MedLens Drug-Drug Interaction (DDI) Evidence

This dataset is a curated, multi-regional drug-drug interaction (DDI) and adverse drug event (ADE) signal database compiled from US, EU/EEA, and India regulatory and pharmacovigilance sources.

It is structured as a SQLite database intended for deterministic offline medication-safety lookups.

This also has a mobile version to put in mobile applications.

---

# Dataset Details

## Dataset Description

The MedLens DDI Evidence dataset aggregates prioritized DDI/ADE screening signals from five regional source files covering the US, EU/EEA, and India.

Each record links a drug pair to one or more adverse effects, a severity level, interaction category, mechanism/rationale, evidence basis, source URLs, patient risk flags, and population relevance notes.

The database contains:

- **21,810** known interaction pairs
- **105,460** pair-level adverse effect rows
- **162,600** raw DDI signal rows (original source-level records, pre-deduplication)

### Severity Breakdown Across Interaction Pairs

| Severity | Pairs |
|---|---:|
| Major | 16,174 |
| Moderate | 5,523 |
| Minor | 113 |

- **Curated by:** Ashutosh
- **Language(s):** English (drug names are generic/INN)
- **License:** CC BY 4.0

---

## Dataset Sources

- **Repository:** [MedLens on GitHub](https://github.com/ashutosh-MedLens/MedLens) *(update with actual URL)*

---

# Uses

## Direct Use

- Offline/on-device medication safety screening
- Drug-drug interaction lookup for clinical decision support prototypes
- Pharmacovigilance research and signal analysis
- Training or evaluation data for NLP models that classify or explain DDI severity
- Regional comparison of interaction signals (US vs EU/EEA vs India)

---

## Out-of-Scope Use

- **Not a substitute for clinical judgment or a licensed drug interaction database.**
  Signals are screening-level, not patient-specific diagnoses or causality determinations.

- Should not be used to generate or imply patient-specific medical advice.

- Not suitable for production clinical systems without expert validation and regulatory clearance.

- Does not cover all known drug interactions; coverage is limited to the source files below.

---

# Dataset Structure

The SQLite database contains five tables.

---

## `known_interaction`

Deduplicated drug-pair level interaction records.

| Column | Type | Description |
|---|---|---|
| `drug_a`, `drug_b` | TEXT | Normalized generic drug names (alphabetical order) |
| `severity` | TEXT | Consensus severity: `Major`, `Moderate`, or `Minor` |
| `severity_rank` | INTEGER | Numeric rank (higher = more severe) |
| `row_count` | INTEGER | Number of raw signals contributing to this pair |
| `source_regions_json` | TEXT | JSON array of regions (e.g. `["us","eu/eea"]`) |
| `evidence_bases_json` | TEXT | JSON array of evidence basis strings |
| `source_bases_json` | TEXT | JSON array of source systems |
| `source_urls_json` | TEXT | JSON array of reference URLs |
| `mechanisms_json` | TEXT | JSON array of mechanism/rationale strings |
| `risk_flags_json` | TEXT | JSON array of patient risk flag strings |
| `dataset_types_json` | TEXT | JSON array of dataset type labels |
| `use_case_notes_json` | TEXT | JSON array of use-case notes |

---

## `known_interaction_effect`

One row per `(pair, adverse effect, severity)` combination.

| Column | Type | Description |
|---|---|---|
| `known_interaction_id` | INTEGER | FK → `known_interaction.id` |
| `adverse_effect` | TEXT | Adverse effect name (e.g. `qt prolongation`, `myelosuppression`) |
| `severity` | TEXT | Effect-level severity |
| `row_count` | INTEGER | Number of raw signals with this effect |
| `source_regions_json` | TEXT | JSON array of regions |

---

## `ddi_raw_signal`

Original source-level records before deduplication (**162,600 rows**).

### Key Columns

- `source_file`
- `region`
- `drug1_raw`
- `drug2_raw`
- `normalized_drug1`
- `normalized_drug2`
- `adverse_effect`
- `severity`
- `mechanism_or_rationale`
- `interaction_category`
- `interaction_direction`
- `evidence_basis`
- `source_basis`
- `source_urls`
- `population_relevance`
- `patient_risk_flags`
- `dataset_type`
- `use_case_note`

---

## `evidence_import_file`

Provenance table — one row per source CSV.

| Source File | Region | Rows Seen | Rows Imported | Unique Pairs |
|---|---|---:|---:|---:|
| `usa_prioritized_ddi_ade_signals.csv` | us | 33,306 | 31,666 | 5,770 |
| `eu_eea_prioritized_ddi_ade_signals.csv` | eu/eea | 58,567 | 50,911 | 8,724 |
| `india_prioritized_ddi_ade_signals.csv` | india | 10,430 | 10,139 | 1,915 |
| `india_expanded_prioritized_ddi_ade_signals.csv` | india_expanded | 55,297 | 49,188 | 14,565 |
| `india_common_generic_ddi_5000.csv` | india_common_generic | 5,000 | 5,000 | 3,471 |

---

## `ddi_import_issue`

Contains **15,696 rows** flagging drug name pairs that could not be resolved to canonical entries during import.

This table is useful for improving normalization coverage.

---

# Dataset Creation

## Curation Rationale

Existing open DDI datasets tend to be US-centric or require a network connection.

This dataset was assembled to enable fully offline, multi-regional medication safety checks — particularly for India, where brand-name and generic drug usage patterns differ significantly from the US/EU baseline.

---

# Source Data

## Data Collection and Processing

Raw DDI/ADE signals were compiled from the following public regulatory and pharmacovigilance sources.

---

### US Signals

- [DailyMed](https://dailymed.nlm.nih.gov/dailymed/) — FDA-approved drug labeling
- [TWOSIDES](https://tatonettilab.org/offsides/) — Post-market drug interaction database
- [CMS Part D Drug Spending Dashboard](https://data.cms.gov/tools/medicare-part-d-drug-spending-dashboard)
- ClinCalc Top-300 Drugs 2023

---

### EU/EEA Signals

- EMA SmPC (Summary of Product Characteristics)
- EMA DDI Guideline
- [EudraVigilance](https://www.ema.europa.eu/en/human-regulatory-overview/research-and-development/pharmacovigilance-research-and-development/eudravigilance) — EU adverse reaction reports
- EU ADR Reports / EU Union Register
- WHO ATC/DDD Index

---

### India Signals

- [PVPI](https://ipc.gov.in/PvPI/pv.html) — Pharmacovigilance Programme of India
- NLEM 2022 (National List of Essential Medicines)
- TWOSIDES

Drug names were normalized to generic/INN form.

Pairs were deduplicated by canonical `(drug_a, drug_b)` sorted order.

Severity was aggregated by taking the highest severity signal per pair.

---

## Who Are the Source Data Producers?

Data originates from:

- US FDA (labeling, FAERS)
- European Medicines Agency (EMA)
- Indian Pharmacopoeia Commission / PVPI
- Academic pharmacoepidemiology projects (TWOSIDES)
- CMS (Medicare Part D utilization data)

---

# Personal and Sensitive Information

This dataset contains no patient-level data.

All records are drug-pair-level aggregate signals.

No personally identifiable information (PII) is present.

---

# Bias, Risks, and Limitations

- **Signal, not causality:**  
  All records are screening-level DDI/ADE signals. They indicate a potential interaction of concern, not a confirmed causal relationship for any individual patient.

- **Coverage gaps:**  
  15,696 raw signals (≈9%) could not be resolved to canonical drug names and are excluded from the deduplicated interaction table.

- **Severity aggregation:**  
  Where multiple sources disagree on severity, the highest severity is taken. This may overstate risk for some pairs.

- **Regional bias:**  
  India coverage is strong for common generics but may underrepresent rare drugs. US coverage skews toward high-prescription-volume drugs (Top-300 proxy).

- **Temporal limitations:**  
  Signal databases are periodically updated; this snapshot reflects the state at import time.

- **Not exhaustive:**  
  Absence of a pair from this dataset does not mean the combination is safe.

---

# Recommendations

- Always cross-reference with current product labeling and consult a qualified healthcare professional before making clinical decisions.

- Use the `ddi_import_issue` table to identify drug names that failed normalization and expand coverage over time.

- Treat `Major` severity pairs as requiring clinical review.

- Treat `Minor` severity pairs as informational.

---

# Glossary

| Term | Meaning |
|---|---|
| **DDI** | Drug-Drug Interaction |
| **ADE** | Adverse Drug Event |
| **FAERS** | FDA Adverse Event Reporting System |
| **PVPI** | Pharmacovigilance Programme of India |
| **NLEM** | National List of Essential Medicines (India) |
| **INN** | International Nonproprietary Name (generic drug name) |
| **TWOSIDES** | A database of drug pair adverse effects derived from post-market surveillance data |
| **SmPC** | Summary of Product Characteristics (EU equivalent of US drug labeling) |

---

# Dataset Card Contact

`prog.mishra@gmail.com`

---

## Citation

If you use this dataset, please cite:

```bibtex
@dataset{mishra2026medlens,
  author       = {Ashutosh Mishra},
  title        = {MedLens: Multi-Regional Drug-Drug Interaction (DDI) Evidence Dataset},
  year         = {2026},
  publisher    = {Hugging Face},
  url          = {https://huggingface.co/datasets/ASHu2/medlens},
  version      = {1.0},
  license      = {CC-BY-4.0}
}
```