Stroke recovery phenotyping through network trajectory approaches and graph neural networks

Abstract

Stroke is a leading cause of neurological injury characterized by impairments in multiple neurological domains including cognition, language, sensory and motor functions. Clinical recovery in these domains is tracked using a wide range of measures that may be continuous, ordinal, interval or categorical in nature, which can present challenges for multivariate regression approaches. This has hindered stroke researchers' ability to achieve an integrated picture of the complex time-evolving interactions among symptoms. Here, we use tools from network science and machine learning that are particularly well-suited to extracting underlying patterns in such data, and may assist in prediction of recovery patterns. To demonstrate the utility of this approach, we analyzed data from the NINDS tPA trial using the Trajectory Profile Clustering (TPC) method to identify distinct stroke recovery patterns for 11 different neurological domains at 5 discrete time points. Our analysis identified 3 distinct stroke trajectory profiles that align with clinically relevant stroke syndromes, characterized both by distinct clusters of symptoms, as well as differing degrees of symptom severity. We then validated our approach using graph neural networks to determine how well our model performed predictively for stratifying patients into these trajectory profiles at early vs. later time points post-stroke. We demonstrate that trajectory profile clustering is an effective method for identifying clinically relevant recovery subtypes in multidimensional longitudinal datasets, and for early prediction of symptom progression subtypes in individual patients. This paper is the first work introducing network trajectory approaches for stroke recovery phenotyping, and is aimed at enhancing the translation of such novel computational approaches for practical clinical application.

Keywords: Disease subtyping; Graph neural networks; Network medicine; Network science; Stroke recovery.

Fig. 1

Data from the NINDS tPA trial shown as a function of time and the 15 NIHSS assessment items. Note that some items (e.g., ataxia) range from 0 to 2, others from 0 to 3 (e.g., language), and others from 0 to 4 (e.g., arm and leg measures). The proportion of participant obtaining that score at each timepoint is shown via a color-coded stacked histogram (total N = 489)

Fig. 2

Correlogram showing the association between slopes in the different domains of the NIHSS. Correlations are shown as Spearman rank-order correlations. Red boxes indicate positive correlations and blue colors indicate negative correlations

Fig. 3

Corresponding profiles of the 3 stroke recovery subtypes. Subtypes identified by the algorithm containing fewer than 10 patients are not shown (1 outlier patient falls under this category). The shade of grey indicates the affected fraction, i.e., fraction of patients in the recovery subtype that are severely affected by that symptom at that time. The number of patients in the subtype, and fraction of patients receiving treatment is listed above each panel. The symptom names are listed to the left. The red boxes highlight the unique combination of dominant symptoms of the ‘left motor’ subtype. The blue boxes highlight the unique combination of dominant symptoms of the middle ‘right motor’ subtype. The rightmost ‘mildly affected’ subtype has the mildest symptom profile. The symptoms names on the left are preceded by the letter ‘S’ (indicating ‘Stroke’) to be consistent with the naming convention in the dataset

Fig. 4

Trajectory profiles (same as in the above figure) were applied independently on patients that A received tissue plasminogen activator (tPA) treatment within 3 h of stroke onset compared to B patients who received placebo. Subtypes identified by the algorithm containing fewer than 10 patients are not shown (1 outlier patient falls under this category). The shade of grey indicates the affected fraction

Fig. 5

Test accuracy denoting predictive power of a graph neural network as a function of timepoint. Accuracy plotted separately for patients that received tPA treatment, placebo patients and all patients (tPA treated+placebo). 70% data used for training the neural network, 30% for testing. Number of training epochs = 100. We use a 2 layer graph convolutional neural network with 16 hidden units and relu nonlinearity at both layers