Test In Colab

User Guide Tutorial 05: Time-to-event Analysis

This tutorial shows how to use TemporAI time_to_event plugins.

All time_to_event plugins

Time-to-event analysis, in the context of TemporAI, refers to models that estimate risk over time for each sample. This may also be referred to as survival analysis.

To see all the relevant plugins:

[ ]:

from tempor import plugin_loader

plugin_loader.list()["time_to_event"]

['ts_coxph', 'ts_xgb', 'dynamic_deephit']

Now also load data source(s) we will use:

[ ]:

PBCDataSource = plugin_loader.get_class("time_to_event.pbc", plugin_type="datasource")

Using a time-to-event analysis plugin.

[ ]:

from tempor import plugin_loader

dataset = PBCDataSource(random_state=42).load()
print(dataset)

model = plugin_loader.get("time_to_event.dynamic_deephit", n_iter=50)
print(model)

TimeToEventAnalysisDataset(
    time_series=TimeSeriesSamples([312, *, 14]),
    static=StaticSamples([312, 1]),
    predictive=TimeToEventAnalysisTaskData(targets=EventSamples([312, 1]))
)
DynamicDeepHitTimeToEventAnalysis(
    name='dynamic_deephit',
    category='time_to_event',
    plugin_type='method',
    params={
        'n_iter': 50,
        'batch_size': 100,
        'lr': 0.001,
        'n_layers_hidden': 1,
        'n_units_hidden': 40,
        'split': 100,
        'rnn_mode': 'GRU',
        'alpha': 0.34,
        'beta': 0.27,
        'sigma': 0.21,
        'dropout': 0.06,
        'device': 'cpu',
        'val_size': 0.1,
        'patience': 20,
        'output_mode': 'MLP',
        'random_state': 0
    }
)

[ ]:

# Targets:
dataset.predictive.targets

EventSamples with data:

status
sample_idx
1 (0.569488555470374, True)
2 (14.1523381885883, False)
3 (0.7365020260650499, True)
4 (0.27653050049282957, True)
5 (4.12057824991786, False)
... ...
308 (4.98850071186069, False)
309 (4.55317051801555, False)
310 (4.4025846019056, False)
311 (4.12879202716022, False)
312 (3.98915781404008, False)

312 rows × 1 columns

[ ]:

# Train.
model.fit(dataset)

DynamicDeepHitTimeToEventAnalysis(
    name='dynamic_deephit',
    category='time_to_event',
    plugin_type='method',
    params={
        'n_iter': 50,
        'batch_size': 100,
        'lr': 0.001,
        'n_layers_hidden': 1,
        'n_units_hidden': 40,
        'split': 100,
        'rnn_mode': 'GRU',
        'alpha': 0.34,
        'beta': 0.27,
        'sigma': 0.21,
        'dropout': 0.06,
        'device': 'cpu',
        'val_size': 0.1,
        'patience': 20,
        'output_mode': 'MLP',
        'random_state': 0
    }
)

[ ]:

# Predict:

model.predict(dataset, horizons=[0.25, 0.50, 0.75])

TimeSeriesSamples with data:

risk_score
sample_idx time_idx
1 0.25 0.422638
0.50 0.626617
0.75 0.717204
2 0.25 0.221283
0.50 0.288457
... ... ...
311 0.50 0.039307
0.75 0.048353
312 0.25 0.248848
0.50 0.390068
0.75 0.444567

936 rows × 1 columns

Note:

The current Dynamic DeepHit implementation has the following limitations: - Only one output feature is supported (no competing risks). - Risk prediction for time points beyond the last event time in the dataset may throw errors.