OHDSI Home | Forums | Wiki | Github

HADES development announcements

EmpiricalCalibration 3.1.1 has been released. This version brings nothing new, but meets a requirement from CRAN to not get kicked out of CRAN (that the package can pass R Check even even the packages listed as suggest are not available).

SelfControlledCaseSeries 3.4.0 has been released. Most important update is that this new version works better with the new Cyclops version.

CohortMethod 4.2.3 has been released. This is a patch release to work better with the new Cyclops version.

EvidenceSynthesis 0.3.0 has been released. The most important change is the support for adaptive grid likelihood approximations for Bayesian meta-analysis. Also, with this release we got EvidenceSynthesis back in CRAN after losing it due to Cyclops temporarily being removed.

IterativeHardThresholding v1.0.2 has been released. This package has been accepted to R CRAN, and is part of our OHDSI HADES suite.

This is an R package for performing L0-based regressions using Cyclops. The specific application that motivated this development was to support patient-level prediction, where there is a desire to select a small number of features into the model (e.g. to create a ‘simple model’ that may have benefits in interpretability, but can be compared against performance of full L1 model to understand the discrimination sacrifice that is made). IterativeHardThresholding fits a large-scale regression model with a penalty that restricts the maximum number of non-zero regression coefficients to a prespecified value.

Big thanks to @msuchard for leading this important contribution to our community. Happy hacking everyone!

1 Like

SqlRender 1.10.0 has been released. Most important changes are added support for Snowflake, Synapse, and DuckDB.

DatabaseConnector 5.1.0 has been released. Most important changes are support for Snowflake, and updated RedShift drivers.

CohortGenerator v0.7.0 has been released. This release includes a way to generate negative control outcome cohorts and some other bug fixes.

DeepPatientLevelPrediction v1.0.1 has been released. This is the the first release and this package has been added to Hades.

This is an R package built on top of PatientLevelPrediction but with deep learning models and required functionality for those. With this package we can train deep learning models using PLP and use all the power of that package but for deep learning. Internally it uses a package developed by rstudio called torch in R which is an R wrapper over the c++ engine that the most popular deep learning framework for research uses (pytorch). Currently no python is required to run the package.

This first release includes three different models, a multilayer perceptron (MLP), a residual network (ResNet) and a Transformer. If you are curious and want to try to develop your first deep learning model the best place to start is this tutorial.

I (Egill Fridgeirsson) will be the maintainer but there is a whole team of people that helped make the package. Thanks to @jreps @Chungsoo_Kim @lhjohn and @SCYou

There will be a poster at the upcoming symposium, presented by @Chungsoo_Kim for anyone who is curious how some deep learning models stack against the tried and tested LASSO and XGBoost on a single prediction problem.

We’d appreciate anyone who wants to try the package and help us make it better. And please if you find issues or have questions slide into our issue tracker

Time to join the hype and go deep.


ResultModelManager v0.1.0 has been released.
This is a new package that is intended to aid the management of Result databases exported or created by packages such as CohortDiagnostics.

The goal of this package is to provide a lightweight utility for data migrations allowing old results to work with new shiny apps and data visualisations.

CohortDiagnostics v3.1.0 has been Released.
This release contains a large number of bug fixes and changes, please check the news for more details.

Note: The shiny app of this version maintains backward compatibility with data generated in version 3.0.0 and above. Attempts will be made to maintain this compatibility.


There is a new HADES package - CohortExplorer

The intent of this package is to help with Phenotype Development and Evaluation. It allows you to take a random sample of persons in the instantiated cohort table and review each individual case profile as time stamped observations across the various domains (condition, observation, procedure, drug, visit etc) - in relation to cohort_start_date. I find this package useful to answer targeted questions that can help improve the cohort definition.

Please note: this package will need access to person level data. The package provides procedures to help further de-identify the data.

1 Like

SqlRender 1.11.0 has been released. Most important changes are

  • Various translation rules that will be required by the upcoming DatabaseConnector release to support dbplyr,
  • Translation from OhdsiSql to SQL Server to support older SQL Server versions (<= 2016) that don’t yet support DROP TABLE IF EXISTS. This is a temporary solution.
1 Like

ResultModelManager v0.1.1 has been released.
This release added schema generator function that creates basic schema creation sql from csv ddl files with table definitions, and allows setting of snakeCaseToCamelCase in a ConnectionHandelr object in a persistent way.

CohortDiagnostics v3.1.1 has been released.
This release contains a bug fixes for the shiny application as well as support for custom features in the FeatureExtraction characterization.

The CohortExplorer shiny app has also been removed as it’s now part of a dedicated new package.

PheValuator 2.1.13 has been released. This is the first release of this package that has all the requirements for HADES.

PheValuator is an R package built using many other HADES packages including PatientLevelPrediction. The package is used to determine the performance characteristics, i.e., sensitivity, specificity, and positive and negative predictive value.

PheValuator follows a multi-step process to determine the performance characteristics of phenotype algorithms. In the first step, phenotype algorithms must be developed manually to identify subjects very likely with and those very likely without the health condition of interest. Those algorithms are used to generate highly accurate labels for subjects in the next step of the process, where a diagnostic predictive model is trained for the health condition. The diagnostic predictive model is next applied to a large, random sample of subjects from the database to determine the probability of each of these subjects of having the health condition. We refer to this set of subjects as the “probabilistic gold standard” dataset. In the final step, we evaluate potential research phenotype algorithms against the probabilistic gold standard and, based on the estimated probabilities, determine the performance characteristics of the tested algorithm. PheValuator has been used on many different type of phenotype algorithms, both those for acute health conditions as well as chronic health conditions.


ParallelLogger 3.1.0 has been released. This includes some quality-of-life improvements, and a minor bug fix.

Some what improvements? :slight_smile:

1 Like

OhdsiShinyModules 1.0.1 has been released. This R package contains shiny modules that can be reused inside other shiny apps presenting OHDSI results. The package currently contains modules for:

  • Prediction
  • CohortMethod
  • Characterization
  • Data Diagnostic
  • CohortGenerator

Please see here if you are interested in adding new shiny modules.

DeepPatientLevelPrediction 1.0.2 has been released. This includes a small fix to make sure the dependencies install correctly when running tests from outside the package.