vigipy

vigipy is a Python library bringing modern disproportionality analyses and pharmacovigilance techniques into the Python ecosystem with a clean, intuitive, and type-safe interface. Core disproportionality methods are adapted and extended from Ismail Ahmed and Antoine Poncet's PhViD package, fully vectorized with native NumPy and SciPy routines.

Top-level Functions & Classes:

• Unified Interface:
  • analyze() - Execute any disproportionality analysis via typed configuration dataclasses
  • analyze_all() - Run multiple analysis methods in a single call with shared or distinct parameters
  • PRRConfig, RORConfig, RFETConfig, BCPNNConfig, GPSConfig, LASSOConfig - Type-safe configuration dataclasses
• Disproportionality Methods:
  • prr() - Proportional Reporting Ratio (frequentist log-normal approximation)
  • ror() - Reporting Odds Ratio (Woolf log-odds approximation)
  • rfet() - Reporting Fisher's Exact Test (exact hypergeometric p-values with optional mid-p correction)
  • bcpnn() - Bayesian Confidence Propagation Neural Network (analytical or Dirichlet Monte Carlo Information Component)
  • gps() - Multi-item Gamma Poisson Shrinker (Empirical Bayes bivariate mixture model)
  • lasso() - LASSO regression for multivariate signal detection and confounding adjustment
• Longitudinal Modeling:
  • LongitudinalModel() - Apply any analysis method over time to evaluate cumulative or disjoint signal evolution
• Data Preparation:
  • convert() - Convert adverse event and product count tables into a structured DataContainer
  • convert_binary() - Generate binary product feature matrices and event outcome matrices for LASSO
  • convert_multi_item() - Aggregate co-occurring product columns into multi-item interaction tables
• Result & Data Containers:
  • AnalysisResult - Structured container for signals, all_signals, num_signals, and model params, with .export() to Excel or CSV
  • DataContainer - Typed container holding contingency, event, and product matrices

Getting Started

Dependencies

vigipy requires Python 3.9+ and modern scientific computing libraries:

• pandas>=2.0
• numpy>=1.24,<3
• scipy>=1.10
• scikit-learn>=1.3
• statsmodels>=0.14

Optional dependencies:

• openpyxl>=3.0.0 (required for exporting results directly to Excel .xlsx spreadsheets)

Installation

Install vigipy from source or local checkout:

pip install .

To include Excel export capabilities:

pip install ".[excel]"

For development (includes test and lint tools):

pip install -e ".[dev]"

Running Tests

Run the full test suite using pytest:

pytest test/ -v

Usage

Unified API (Recommended)

The unified interface provides type safety, autocompletion, and consistent result structures across all methods:

import pandas as pd
from vigipy import convert, analyze, analyze_all, PRRConfig, BCPNNConfig, GPSConfig

# 1. Load data and convert to a DataContainer
df = pd.read_csv("AE_count_data.csv")
data = convert(df, product_label="name", ae_label="AE", count_label="count")

# 2. Run a single method with typed configuration
result = analyze(data, PRRConfig(min_events=3, decision_metric="fdr", fdr_threshold=0.05))

print(f"Detected {result.num_signals} signals")
print(result.signals.head())

# Export both significant signals and full dataset to Excel or CSV
result.export("prr_signals.xlsx")   # Creates 'Signals' and 'all_data' sheets
result.export("prr_signals.csv")    # Exports detected signals to CSV

# 3. Batch comparison across all methods in one call
batch_results = analyze_all(data, min_events=3, decision_metric="rank")
for method_name, res in batch_results.items():
    print(f"{method_name.upper()}: {res.num_signals} signals detected")

# 4. Iterate over custom configurations
configs = [
    PRRConfig(min_events=5, ranking_statistic="CI"),
    BCPNNConfig(min_events=5, ranking_statistic="quantile"),
    GPSConfig(min_events=5, ranking_statistic="log2"),
]
for cfg in configs:
    res = analyze(data, cfg)
    print(f"{cfg.method}: {res.num_signals} signals")

Classic Function API

Direct function calls are fully supported with identical return structures:

import pandas as pd
from vigipy import convert, prr, ror, rfet, bcpnn, gps

df = pd.read_csv("AE_count_data.csv")
data = convert(df)

# Frequentist: Reporting Odds Ratio with Haldane-Anscombe continuity correction
ror_res = ror(data, min_events=3, decision_metric="fdr", fdr_threshold=0.05)

# Fisher's Exact Test with Lancaster mid-p adjustment
rfet_res = rfet(data, min_events=3, mid_pval=True)

# Bayesian Confidence Propagation Neural Network
bcpnn_res = bcpnn(data, min_events=3, ranking_statistic="quantile")

# Empirical Bayes: Gamma Poisson Shrinker
gps_res = gps(data, min_events=5, decision_metric="rank", ranking_statistic="log2")

# Access results
print(gps_res.signals[["Product", "Adverse Event", "Count", "quantile", "fdr"]].head())
gps_res.export("gps_signals.xlsx")

Decision Rules & Ranking Statistics

vigipy standardizes signal identification across frequentist and Bayesian methods:

Decision Metrics (decision_metric)

• "fdr" - Controls the False Discovery Rate at decision_thres (default: 0.05) using Local Bayes Estimation (LBE) or cumulative posterior null probabilities.
• "rank" - Retains signals where the ranking statistic meets decision_thres. For p-values, selects values $\le \text{threshold}$; for confidence/credible bounds, selects values $\ge \text{threshold}$.
• "signals" - Selects the top $N$ ranked candidates up to decision_thres.

Ranking Statistics (ranking_statistic)

Expected Count Calculations & Dispersion Testing

Expected counts ($E$) model the baseline event frequency under the null hypothesis of no association. vigipy supports three expectation models:

1. "mantel-haentzel" (default): Standard independence assumption, $E_{ij} = \frac{n_{i\cdot} n_{\cdot j}}{N}$.
2. "poisson": Generalized Linear Model using Poisson log-linear regression.
3. "negative-binomial": Generalized Linear Model with negative binomial dispersion parameter method_alpha.

When event data exhibits overdispersion (variance significantly exceeds the mean), Poisson estimates may underestimate variance. You can test for overdispersion using Cameron and Trivedi's auxiliary regression test:

import pandas as pd
from vigipy import convert, bcpnn
from vigipy.utils import test_dispersion

df = pd.read_csv("AE_count_data.csv")
data = convert(df)

# Test for overdispersion
dispersion_info = test_dispersion(data)
print(f"Dispersion ratio: {dispersion_info['dispersion']:.2f}")

# If overdispersed (> 2), use the estimated alpha in Negative Binomial regression
alpha = dispersion_info["alpha"] if dispersion_info["dispersion"] > 2 else 1.0
res = bcpnn(data, expected_method="negative-binomial", method_alpha=alpha, min_events=3)

Longitudinal Modeling

The LongitudinalModel class evaluates disproportionality over time to monitor signal emergence, stability, and trajectory.

You can run models in two modes:

• Cumulative (run): Progressively incorporates historical data up to each resampled time boundary, tracking accumulating evidence.
• Disjoint (run_disjoint): Evaluates each time window independently without historical accumulation.

import pandas as pd
from vigipy import LongitudinalModel, gps

df = pd.read_csv("AE_time_series.csv")
# Must contain: 'date', 'name', 'AE', and 'count' (or custom count_col)

# Initialize grouped by calendar year ('YE', 'QE', 'ME')
lm = LongitudinalModel(df, time_unit="YE", count_col="count")

# Run GPS cumulatively over time
lm.run(gps, include_gaps=False, decision_metric="rank", ranking_statistic="log2")

# Regroup to quarterly slices and evaluate
lm.regroup_dates("QE")
lm.run_disjoint(gps, include_gaps=False, decision_metric="rank", ranking_statistic="log2")

# Access results chronologically: (timestamp, AnalysisResult)
for timestamp, result in lm.results:
    if result is not None:
        print(f"Slice ending {timestamp.date()}: {result.num_signals} signals")
        print(result.signals.head(2))

LASSO Signal Detection

LASSO regression models multiple products simultaneously, adjusting for co-prescriptions, confounding by indication, and polypharmacy.

1. Pure Binary Matrix

Each row represents an individual report or subject:

import pandas as pd
from vigipy import convert_binary, lasso

df = pd.read_csv("patient_reports.csv")
bin_data = convert_binary(df, product_label="name", ae_label="AE", use_counts=False)

# Linear LASSO with Information Criterion model selection
result = lasso(bin_data, use_IC=True, IC_criterion="bic", min_events=3)
print(result.signals[["Product", "Adverse Event", "LASSO Coefficient", "CI Lower", "CI Upper"]])
result.export("lasso_signals.xlsx")

2. Count Outcomes & GLM LASSO

When aggregate event counts are used:

bin_data = convert_binary(df, product_label="name", ae_label="AE", use_counts=True)

# Fit Negative Binomial GLM with L1 regularization
result = lasso(bin_data, use_glm=True, lasso_thresh=0.2, nb_alpha=1.0)
result.export("lasso_glm_signals.csv")

Multi-Item Interaction Conversion

To analyze interactions between co-occurring drugs or devices:

from vigipy.utils.data_prep import convert_multi_item

# Aggregate co-administered products
multi_data = convert_multi_item(
    df,
    product_label=["suspect_drug_1", "suspect_drug_2"],
    ae_label="AE",
    count_label="count",
    min_threshold=3,
)

The returned DataContainer is fully compatible with prr, ror, rfet, bcpnn, and gps.

Result Inspection & Export

All analysis methods return an AnalysisResult object:

result = analyze(data, PRRConfig(min_events=3))

# Filtered signals meeting decision criteria
signals_df = result.signals

# All evaluated candidate pairs with computed statistics
all_df = result.all_signals

# Number of identified signals
count = result.num_signals

# Input parameters and model metadata
params_dict = result.params

# Export to Excel (.xlsx) or CSV (.csv)
result.export("output.xlsx")  # Writes 'Signals' and 'all_data' sheets
result.export("output.csv")   # Writes signals DataFrame

Authors

• David Beery (@Shakesbeery)

License

This project is licensed under the MIT License.

Acknowledgements

• Ismail Ahmed and Antoine Poncet for the foundational design of the PhViD package in R.
• Ross Ihaka and Catherine Loader for early mathematical formulations of deviance and log-gamma approximations.