TemporAI

⚗️ Status: This project is still in alpha, and the API may change without warning.

📃 Overview

TemporAI is a Machine Learning-centric time-series library for medicine. The tasks that are currently of focus in TemporAI are: time-to-event (survival) analysis with time-series data, treatment effects (causal inference) over time, and time-series prediction. Data preprocessing methods, including missing value imputation for static and temporal covariates, are provided. AutoML tools for hyperparameter tuning and pipeline selection are also available.

How is TemporAI unique?

• 🏥 Medicine-first: Focused on use cases for medicine and healthcare, such as temporal treatment effects, survival analysis over time, imputation methods, models with built-in and post-hoc interpretability, ... See methods.
• 🏗️ Fast prototyping: A plugin design allowing for on-the-fly integration of new methods by the users.
• 🚀 From research to practice: Relevant novel models from research community adapted for practical use.
• 🌍 A healthcare ecosystem vision: A range of interactive demonstration apps, new medical problem settings, interpretability tools, data-centric tools etc. are planned.

Key concepts

🚀 Installation

Instal with pip

From the Python Package Index (PyPI):

$ pip install temporai

Or from source:

$ git clone https://github.com/vanderschaarlab/temporai.git
$ cd temporai
$ pip install .

Install in a conda environment

While have not yet published TemporAI on conda-forge, you can still install TemporAI in your conda environment using pip as follows:

Create and activate conda environment as normal:

$ conda create -n <my_environment>
$ conda activate <my_environment>

Then install inside your conda environment with pip:

$ pip install temporai

💥 Sample Usage

(▶️ Expand to view the sections below.)

List the available plugins

from tempor import plugin_loader

print(plugin_loader.list())

Use a time-to-event (survival) analysis model

from tempor import plugin_loader

# Load a time-to-event dataset:
dataset = plugin_loader.get("time_to_event.pbc", plugin_type="datasource").load()

# Initialize the model:
model = plugin_loader.get("time_to_event.dynamic_deephit")

# Train:
model.fit(dataset)

# Make risk predictions:
prediction = model.predict(dataset, horizons=[0.25, 0.50, 0.75])

Use a temporal treatment effects model

import numpy as np

from tempor import plugin_loader

# Load a dataset with temporal treatments and outcomes:
dataset = plugin_loader.get(
    "treatments.temporal.dummy_treatments",
    plugin_type="datasource",
    temporal_covariates_missing_prob=0.0,
    temporal_treatments_n_features=1,
    temporal_treatments_n_categories=2,
).load()

# Initialize the model:
model = plugin_loader.get("treatments.temporal.regression.crn_regressor", epochs=20)

# Train:
model.fit(dataset)

# Define target variable horizons for each sample:
horizons = [
    tc.time_indexes()[0][len(tc.time_indexes()[0]) // 2 :] for tc in dataset.time_series
]

# Define treatment scenarios for each sample:
treatment_scenarios = [
    [np.asarray([1] * len(h)), np.asarray([0] * len(h))] for h in horizons
]

# Predict counterfactuals:
counterfactuals = model.predict_counterfactuals(
    dataset,
    horizons=horizons,
    treatment_scenarios=treatment_scenarios,
)

Use a missing data imputer

from tempor import plugin_loader

dataset = plugin_loader.get(
    "prediction.one_off.sine", plugin_type="datasource", with_missing=True
).load()
static_data_n_missing = dataset.static.dataframe().isna().sum().sum()
temporal_data_n_missing = dataset.time_series.dataframe().isna().sum().sum()

print(static_data_n_missing, temporal_data_n_missing)
assert static_data_n_missing > 0
assert temporal_data_n_missing > 0

# Initialize the model:
model = plugin_loader.get("preprocessing.imputation.temporal.bfill")

# Train:
model.fit(dataset)

# Impute:
imputed = model.transform(dataset)
temporal_data_n_missing = imputed.time_series.dataframe().isna().sum().sum()

print(static_data_n_missing, temporal_data_n_missing)
assert temporal_data_n_missing == 0

Use a one-off classifier (prediction)

from tempor import plugin_loader

dataset = plugin_loader.get("prediction.one_off.sine", plugin_type="datasource").load()

# Initialize the model:
model = plugin_loader.get("prediction.one_off.classification.nn_classifier", n_iter=50)

# Train:
model.fit(dataset)

# Predict:
prediction = model.predict(dataset)

Use a temporal regressor (forecasting)

from tempor import plugin_loader

# Load a dataset with temporal targets.
dataset = plugin_loader.get(
    "prediction.temporal.dummy_prediction",
    plugin_type="datasource",
    temporal_covariates_missing_prob=0.0,
).load()

# Initialize the model:
model = plugin_loader.get("prediction.temporal.regression.seq2seq_regressor", epochs=10)

# Train:
model.fit(dataset)

# Predict:
prediction = model.predict(dataset, n_future_steps=5)

Benchmark models, time-to-event task

from tempor.benchmarks import benchmark_models
from tempor import plugin_loader
from tempor.methods.pipeline import pipeline

testcases = [
    (
        "pipeline1",
        pipeline(
            [
                "preprocessing.scaling.temporal.ts_minmax_scaler",
                "time_to_event.dynamic_deephit",
            ]
        )({"ts_coxph": {"n_iter": 100}}),
    ),
    (
        "plugin1",
        plugin_loader.get("time_to_event.dynamic_deephit", n_iter=100),
    ),
    (
        "plugin2",
        plugin_loader.get("time_to_event.ts_coxph", n_iter=100),
    ),
]
dataset = plugin_loader.get("time_to_event.pbc", plugin_type="datasource").load()

aggr_score, per_test_score = benchmark_models(
    task_type="time_to_event",
    tests=testcases,
    data=dataset,
    n_splits=2,
    random_state=0,
    horizons=[2.0, 4.0, 6.0],
)

print(aggr_score)

Serialization

from tempor.utils.serialization import load, save
from tempor import plugin_loader

# Initialize the model:
model = plugin_loader.get("prediction.one_off.classification.nn_classifier", n_iter=50)

buff = save(model)  # Save model to bytes.
reloaded = load(buff)  # Reload model.

# `save_to_file`, `load_from_file` also available in the serialization module.

AutoML - search for the best pipeline for your task

from tempor.automl.seeker import PipelineSeeker

dataset = plugin_loader.get("prediction.one_off.sine", plugin_type="datasource").load()

# Specify the AutoML pipeline seeker for the task of your choice, providing candidate methods,
# metric, preprocessing steps etc.
seeker = PipelineSeeker(
    study_name="my_automl_study",
    task_type="prediction.one_off.classification",
    estimator_names=[
        "cde_classifier",
        "ode_classifier",
        "nn_classifier",
    ],
    metric="aucroc",
    dataset=dataset,
    return_top_k=3,
    num_iter=100,
    tuner_type="bayesian",
    static_imputers=["static_tabular_imputer"],
    static_scalers=[],
    temporal_imputers=["ffill", "bfill"],
    temporal_scalers=["ts_minmax_scaler"],
)

# The search will return the best pipelines.
best_pipelines, best_scores = seeker.search()  # doctest: +SKIP

📖 Tutorials

Data

• - Data Format
• - Datasets
• - Data Loaders
• - Data Splitting
• - Other Data Formats
• - MIMIC Use Case

User Guide

• - Plugins
• - Imputation
• - Scaling
• - Prediction
• - Time-to-event Analysis
• - Treatment Effects
• - Pipeline
• - Benchmarks
• - AutoML

Extending TemporAI

• - Writing a Custom Method Plugin
• - Testing a Custom Method Plugin
• - Writing a Custom Data Source Plugin
• - Writing a Custom Metric Plugin
• - Writing a Custom Data Format

📘 Documentation

See the full project documentation here.

Note on documentation versions:

• If you have installed TemporAI from PyPI, you should refer to the stable documentation.
• If you have installed TemporAI from source, you should refer to the latest documentation.

See the Instal with pip section for reference.

🌍 TemporAI Ecosystem (Experimental)

We provide additional tools in the TemporAI ecosystem, which are in active development, and are currently (very) experimental. Suggestions and contributions are welcome!

These include:

• temporai-clinic: A web app tool for interacting and visualising TemporAI models, data, and predictions.
• temporai-mivdp: A MIMIC-IV-Data-Pipeline adaptation for TemporAI.

🔑 Methods

(▶️ Expand to view the sections below.)

Time-to-Event (survival) analysis over time

Risk estimation given event data (category: time_to_event)

Name	Description	Reference
dynamic_deephit	Dynamic-DeepHit incorporates the available longitudinal data comprising various repeated measurements (rather than only the last available measurements) in order to issue dynamically updated survival predictions	Paper
ts_coxph	Create embeddings from the time series and use a CoxPH model for predicting the survival function	---
ts_xgb	Create embeddings from the time series and use a SurvivalXGBoost model for predicting the survival function	---

Treatment effects

One-off

Treatment effects estimation where treatments are a one-off event.

• Regression on the outcomes (category: treatments.one_off.regression)

Name	Description	Reference
synctwin_regressor	SyncTwin is a treatment effect estimation method tailored for observational studies with longitudinal data, applied to the LIP setting: Longitudinal, Irregular and Point treatment.	Paper

Temporal

Treatment effects estimation where treatments are temporal (time series).

• Classification on the outcomes (category: treatments.temporal.classification)

Name	Description	Reference
crn_classifier	The Counterfactual Recurrent Network (CRN), a sequence-to-sequence model that leverages the available patient observational data to estimate treatment effects over time.	Paper

• Regression on the outcomes (category: treatments.temporal.regression)

Name	Description	Reference
crn_regressor	The Counterfactual Recurrent Network (CRN), a sequence-to-sequence model that leverages the available patient observational data to estimate treatment effects over time.	Paper

Prediction

One-off

Prediction where targets are static.

• Classification (category: prediction.one_off.classification)

Name	Description	Reference
nn_classifier	Neural-net based classifier. Supports multiple recurrent models, like RNN, LSTM, Transformer etc.	---
ode_classifier	Classifier based on ordinary differential equation (ODE) solvers.	---
cde_classifier	Classifier based Neural Controlled Differential Equations for Irregular Time Series.	Paper
laplace_ode_classifier	Classifier based Inverse Laplace Transform (ILT) algorithms implemented in PyTorch.	Paper

• Regression (category: prediction.one_off.regression)

Name	Description	Reference
nn_regressor	Neural-net based regressor. Supports multiple recurrent models, like RNN, LSTM, Transformer etc.	---
ode_regressor	Regressor based on ordinary differential equation (ODE) solvers.	---
cde_regressor	Regressor based Neural Controlled Differential Equations for Irregular Time Series.	Paper
laplace_ode_regressor	Regressor based Inverse Laplace Transform (ILT) algorithms implemented in PyTorch.	Paper

Temporal

Prediction where targets are temporal (time series).

• Classification (category: prediction.temporal.classification)

Name	Description	Reference
seq2seq_classifier	Seq2Seq prediction, classification	---

• Regression (category: prediction.temporal.regression)

Name	Description	Reference
seq2seq_regressor	Seq2Seq prediction, regression	---

Preprocessing

Feature Encoding

• Static data (category: preprocessing.encoding.static)

Name	Description	Reference
static_onehot_encoder	One-hot encode categorical static features	---

• Temporal data (category: preprocessing.encoding.temporal)

Name	Description	Reference
ts_onehot_encoder	One-hot encode categorical time series features	---

Imputation

• Static data (category: preprocessing.imputation.static)

Name	Description	Reference
static_tabular_imputer	Use any method from HyperImpute (HyperImpute, Mean, Median, Most-frequent, MissForest, ICE, MICE, SoftImpute, EM, Sinkhorn, GAIN, MIRACLE, MIWAE) to impute the static data	Paper

• Temporal data (category: preprocessing.imputation.temporal)

Name	Description	Reference
ffill	Propagate last valid observation forward to next valid	---
bfill	Use next valid observation to fill gap	---
ts_tabular_imputer	Use any method from HyperImpute (HyperImpute, Mean, Median, Most-frequent, MissForest, ICE, MICE, SoftImpute, EM, Sinkhorn, GAIN, MIRACLE, MIWAE) to impute the time series data	Paper

Scaling

• Static data (category: preprocessing.scaling.static)

Name	Description	Reference
static_standard_scaler	Scale the static features using a StandardScaler	---
static_minmax_scaler	Scale the static features using a MinMaxScaler	---

• Temporal data (category: preprocessing.scaling.temporal)

Name	Description	Reference
ts_standard_scaler	Scale the temporal features using a StandardScaler	---
ts_minmax_scaler	Scale the temporal features using a MinMaxScaler	---

🔨 Tests and Development

Install the testing dependencies using:

pip install .[testing]

The tests can be executed using:

pytest -vsx

For local development, we recommend that you should install the [dev] extra, which includes [testing] and some additional dependencies:

pip install .[dev]

For development and contribution to TemporAI, see:

• 📓 Extending TemporAI tutorials
• 📃 Contribution guide
• 👩‍💻 Developer's guide

✍️ Citing

If you use this code, please cite the associated paper:

@article{saveliev2023temporai,
  title={TemporAI: Facilitating Machine Learning Innovation in Time Domain Tasks for Medicine},
  author={Saveliev, Evgeny S and van der Schaar, Mihaela},
  journal={arXiv preprint arXiv:2301.12260},
  year={2023}
}