Classify ALS diagnosis from white matter features¶
This example demonstrates classification, a machine learning task in which a model is constructed and fit to accurately discriminate between different classes of subjects. In this case, participants with amyotrophic lateral sclerosis (ALS) and healthy control. We use the dataset from (Sarica et al, 2017), which contains tractometry features from 24 patients with ALS and 24 demographically matched control subjects. We will use the sparse group lasso (SGL) algorithm implemented in AFQ-Insight, as also described in Richie-Halford et al. 2021.
import os.path as op
from paths import afq_home
import matplotlib.pyplot as plt
import numpy as np
from groupyr.decomposition import GroupPCA
from sklearn.impute import SimpleImputer
from sklearn.model_selection import cross_validate
from afqinsight import load_afq_data, make_afq_classifier_pipeline
Load the data from Sarica et al.¶
The data is read into an AFQData object, which manages the division of subject information and outputs of pyAFQ (in the nodes.csv file) and merges between them.
afqdata = load_afq_data(op.join(afq_home, "afq-insight/sarica/nodes.csv"),
op.join(afq_home, "afq-insight/sarica/subjects.csv"),
dwi_metrics=["fa", "md"],
target_cols=["class"],
label_encode_cols=["class"])
/Users/john/AFQ-Insight/afqinsight/transform.py:144: FutureWarning: The previous implementation of stack is deprecated and will be removed in a future version of pandas. See the What's New notes for pandas 2.1.0 for details. Specify future_stack=True to adopt the new implementation and silence this warning. features = interpolated.stack(["subjectID", "tractID", "metric"]).unstack(
Examine the data¶
afqdata is an AFQDataset object, with properties corresponding to
the tractometry features and phenotypic targets.
X = afqdata.X
y = afqdata.y.astype(float) # SGL expects float targets
groups = afqdata.groups
feature_names = afqdata.feature_names
group_names = afqdata.group_names
subjects = afqdata.subjects
Reduce data dimensionality¶
To save computational time, we take the first 10 principal components from each feature group (i.e. from each metric-bundle combination). For more details on this approach in a research setting, please see [2]_.
Here we reduce computation time by taking the first 10 principal components of
each feature group and performing SGL logistic regression on those components.
If you want to train an SGL model without group PCA, set do_group_pca = False. This will increase the number of features by an order of magnitude
and slow down execution time.
do_group_pca = True
if do_group_pca:
n_components = 10
# The next three lines retrieve the group structure of the group-wise PCA
# and store it in ``groups_pca``. We do not use the imputer or GroupPCA transformer
# for anything else
imputer = SimpleImputer(strategy="median")
gpca = GroupPCA(n_components=n_components, groups=groups)
groups_pca = gpca.fit(imputer.fit_transform(X)).groups_out_
transformer = GroupPCA
transformer_kwargs = {"groups": groups, "n_components": n_components}
else:
transformer = False
transformer_kwargs = None
Create the classification pipeline¶
The core computational machinery is a pipeline. These operate as scikit-learn compatible pipelines, so we can pass them to scikit-learn functions. There are many options that need to be set to configure the pipeline object.
pipe = make_afq_classifier_pipeline(
imputer_kwargs={"strategy": "median"}, # Use median imputation
use_cv_estimator=True, # Automatically determine the best hyperparameters
feature_transformer=transformer, # See note above about group PCA
feature_transformer_kwargs=transformer_kwargs,
scaler="standard", # Standard scale the features before regression
groups=(
groups_pca if do_group_pca else groups
), # SGL will use the original feature groups or the PCA feature groups depending on the choice above # noqa E501
verbose=0, # Be quiet!
pipeline_verbosity=False, # No really, be quiet!
tuning_strategy="bayes", # Use BayesSearchCV to determine optimal hyperparameters
n_bayes_iter=20, # Consider only this many points in hyperparameter space
cv=3, # Use three CV splits to evaluate each hyperparameter combination
l1_ratio=[0.0, 1.0], # Explore the entire range of ``l1_ratio``
eps=5e-2, # This is the ratio of the smallest to largest ``alpha`` value
tol=1e-2, # Set a lenient convergence tolerance just for this example
)
Fit and cross-validate¶
The pipe object is a scikit-learn pipeline and can be used in other
scikit-learn functions. Here, the generic cross_validate function.
scores = cross_validate(
pipe, X, y, cv=5, return_train_score=True, return_estimator=True
)
Display results¶
Finally, we display the results, including both the training and test scores, as well as a visualization of the model coefficients: which tracts contributed the most to this classification.
print(f"Mean train score: {np.mean(scores['train_score']):5.3f}")
print(f"Mean test score: {np.mean(scores['test_score']):5.3f}")
print(f"Mean fit time: {np.mean(scores['fit_time']):5.2f}s")
print(f"Mean score time: {np.mean(scores['score_time']):5.2f}s")
mean_coefs = np.mean(
np.abs([est.named_steps["estimate"].coef_ for est in scores["estimator"]]), axis=0
)
fig, ax = plt.subplots(1, 1, figsize=(8, 5))
_ = ax.plot(mean_coefs[:180], color="black", lw=2)
_ = ax.set_xlim(0, 180)
colors = plt.get_cmap("tab20").colors
for grp, grp_name, color in zip(groups_pca[:18], group_names, colors):
_ = ax.axvspan(grp.min(), grp.max() + 1, color=color, alpha=0.8, label=grp_name[1])
box = ax.get_position()
_ = ax.set_position(
[box.x0, box.y0 + box.height * 0.375, box.width, box.height * 0.625]
)
_ = ax.legend(loc="upper center", bbox_to_anchor=(0.5, -0.2), ncol=3)
_ = ax.set_ylabel(r"$\hat{\beta}$", fontsize=16)
_ = ax.set_xlabel("Group principal component", fontsize=16)
_ = ax.set_title("Group Principal Regression Coefficients (FA only)", fontsize=18)
Mean train score: 1.000 Mean test score: 0.791 Mean fit time: 3.22s Mean score time: 0.00s
