Annals of Neurology - uses 24-hour post-treatment features to predict the prognosis after acute cerebral infarction

Over the past decade, machine learning (ML) techniques have been successfully applied to the field of
acute ischemic stroke (AIS).
Current applications include ML algorithms that assist in stroke diagnosis, several of which are commercial, that can help detect large vessel occlusion (LVO) and identify patients
with mismatched perfusion imaging.

Figure 1: Cover image of the paper

Since ML constitutes a diverse category of methods and has the ability to handle linear and nonlinear interactions in datasets, the application of ML algorithms in AIS may be promising
in predicting patient outcomes.
Previous studies used ML models to predict outcomes for AIS patients after MT, essentially reporting similar prediction accuracy
between traditional logistic regression methods and ML algorithms.
Although ML models were trained on both baseline and treatment variables, the methods of variable selection and the availability of variables varied in these studies
.

Some studies have shown that early neurological states (as measured by a 24-hour NIHSS) are associated
with 90-day functional outcomes.
Therefore, the 24-hour NIHSS score remains a potential surrogate marker
of long-term functional outcomes in AIS patients treated with EVT.
Recently, Mistry et al.
studied the 24-hour NIHSS score as a predictor of 90-day functional outcomes in patients in the Multi-Medium Cardiovascular Stroke Treatment (Best) study
.
When adjusted for baseline NIHSS, the 24-hour NIHSS score proved to be the strongest predictor of dichotomous and ordinal 90-day functional outcomes
.

Thus, Alicia C.
Castonguay et al.
, University of Toledo in the United States, studied the predictive accuracy
of multiple ML algorithms trained on the basis of 24-hour NIHSS scores and traditional logistic regression (T-LR) in the Prospective Systematic Evaluation of Neurothrombectomy Device Treatment (STRATIS) registry in patients with acute ischemic stroke.

They used ML models, adaptive boosting, random forests (RF), classification and regression trees (CART), C5.
0 decision trees (C5.
0), support vector machines (SVMs), minimum absolute shrinkage and selection operators (LASSO), and logistic regression (LR), as well as traditional LR models to predict 90-day functional outcomes (modified Rankin scale score 0-2).

In all models, the 24-hour National Institutes of Health Stroke Scale (NIHSS) was studied
as a continuous or dichotomous variable.
Use the area under the characteristic curve (AUC) to assess the accuracy of
the model.

Figure 2: Graph of paper results

In all models, a 24-hour NIHSS score was the primary predictor of
functional outcomes.
ML models using consecutive 24-hour NIHSS scores showed moderate to good predictive performance (mean AUC range: 0.
76-0.
92); however, RF (AUC: 0.
92±0.
028) outperformed all ML models except LASSO (AUC: 0.
89±0.
023, P=0.
0958).

Importantly, when utilizing 24-hour continuous NIHSS scores, RF showed significantly higher predictive values
than LR (AUC: 0.
87±0.
031, p=0.
048) and conventional LR (AUC: 85±0.
06, p=0.
035).
The dichotomy of the 24-hour NIHSS score is similar
to the prediction accuracy of the continuous ML model.

In this substudy, they found similar
prediction accuracy for functional outcomes when using 24-hour NIHSS scores as a continuous or dichotomous variable in ML models.
ML models have moderate to good prediction accuracy, with RF models outperforming LR models
.
External validation
of these ML models is necessary.

Original:
Schirinzi T, Maftei D, Passali FM, et al.
Olfactory neuron Prokineticin‐2 as a potential target in Parkinson's disease.
_Annals of Neurology_.
Published online October 11, 2022:ana.
26526.
doi:[10.
1002/ana.
26526](https://doi.
org/10.
1002/ana.
26526)