Rapid Computer‐Aided Diagnosis of Stroke by Serum Metabolic Fingerprint Based Multi‐Modal Recognition

Abstract Stroke is a leading cause of mortality and disability worldwide, expected to result in 61 million disability‐adjusted life‐years in 2020. Rapid diagnostics is the core of stroke management for early prevention and medical treatment. Serum metabolic fingerprints (SMFs) reflect underlying disease progression, predictive of patient phenotypes. Deep learning (DL) encoding SMFs with clinical indexes outperforms single biomarkers, while posing challenges with poor prediction to interpret by feature selection. Herein, rapid computer‐aided diagnosis of stroke is performed using SMF based multi‐modal recognition by DL, to combine adaptive machine learning with a novel feature selection approach. SMFs are extracted by nano‐assisted laser desorption/ionization mass spectrometry (LDI MS), consuming 100 nL of serum in seconds. A multi‐modal recognition is constructed by integrating SMFs and clinical indexes with an enhanced area under curve (AUC) up to 0.845 for stroke screening, compared to single‐modal diagnosis by only SMFs or clinical indexes. The prediction of DL is addressed by selecting 20 key metabolite features with differential regulation through a saliency map approach, shedding light on the molecular mechanisms in stroke. The approach highlights the emerging role of DL in precision medicine and suggests an expanding utility for computational analysis of SMFs in stroke screening.

Deep learning (DL) method: The original input of both stroke network (SN) and clinical stroke network (CSN) was unified with 1024 dimensional input (x_input), in which 881 features for single-modal SMFs recognition including 881 m/z signals (x_spectral), and 905 features for SMFs based multi-modal recognition including 881 m/z signals (x_spectral) and 24 clinical indexes (x_ext). The rest features (143 for SMF based single-modal recognition and 119 for SMF based multi-modal recognition) were set as 0, to enhance the adaptability of the networks.
All input features were centralized and scaled to (-1,1). The networks were designed based on the deep neural networks (DNN), with two main networks (feature extraction part (feature_extract) and non-linear feature interaction layer (feature_interaction). The reorganized features (96 both for SN and CSN) after extraction and interaction were input to the classification layer (Softmax) for classification probability output.
The detailed principle formula from the original 1024 dimensional input (x_input) to the softmax layer was as follows: The feature extraction part (feature_extract) was consisted of four locally connected 1D layers. [4] Each layer was subdivided into 32 regions for feature extraction. The four-layer feature extraction part was defined as: where 1 , … , 32 represented 32 extracted regions in the l (l = 0, 1, 2, 3) layer, and 1 , … , 32 represented parameters based on Adam optimization algorithm in each 32 regions of discovery subject.
The non-linear feature interaction layer (feature_interaction) was used to analyze the nonlinear relationship among input features. In particular, the residuals from previous layer, discrete relation (ReLU activation), quadratic relation, and non-linear relation (Tanh and sigmoid) were used for linear and non-linear feature transformation. [5] The non-linear feature interaction layer was defined as: where represented features after linear and non-linear feature transformation in the l-1 (l = 1, 2, 3, 4, ……) layer; represented the parameter of ReLU activation; 1 , 2 , represented the parameter of quadratic relation; and represented the parameter of Tanh and sigmoid respectively. The DL framework was constructed using Keras 2.3.1 + tensorflow1.14.0 (GPU acceleration with one Nvidia RTX 2080Ti graphical card).
Classification was performed using a 10-fold cross-validation (repeated for 20 rounds) to assess the diagnostic accuracy within the discovery cohorts. For the permutation test, we randomly permuted disease label ("0" for healthy controls and "1" for patients, previously) for 1000 times and calculated the distribution of AUC using the uninformative data obtained by random permutation.
Machine learning (ML) method: All ML in this work were performed based on Python 3.6 (Anaconda distribution). The least absolute shrinkage and selector operator (LASSO), random forest (RF) [6] and support vector machine (SVM) [6b, 7] were implemented in Python Scikit-learn 0.22.2. For orthogonal partial least squares discriminant analysis (OPLS-DA), [8] the algorithm was derived from PLS-DA algorithms, for the purpose of improving model interpretation and reducing model complexity. For valid comparison, we performed 20 rounds of 10-fold crossvalidation for all ML methods, strictly following the same experimental configuration with the preconstructed DL method.
Computer-assisted diagnosis: The performance of classification models constructed by both DL and ML was measured by sensitivity, specificity, and area under curve (AUC). Receiver operating characteristic (ROC) curves were generated using classification probabilities of stroke patients versus healthy controls and the true labels of each test mass spectrum. The final AUC was an average of all the AUCs obtained from all folds of the cross-validation. Sensitivity and specificity were determined by dividing the total number of correctly labeled stroke patients and the total number of correctly labeled healthy controls, respectively, by the total number of all samples. Notably, the validation cohort was independent of the discovery cohort in the diagnosis stage for blind testing.
Biomarker selection: The significant variables by DL were identified and selected by activation maximization (AM). Typically, each SMF (mass spectrum) was converted to a twodimensional graphic. [9] All graphics were listed in sequence, shown as a saliency map ( Figure   S5) with discriminant significance for each input feature towards feature selection. The conversion was mainly dependent on Keras Visualization Toolkit (https://github.com/raghakot/keras-vis/blob/master/README.md). Top 20 m/z signals were selected with score over 0.2 (the discriminant saliency derived from saliency map), contributing the most to the differentiation between healthy controls and stroke patients. The as-selected 20 m/z signals were validated as metabolite feature panel by accurate mass measurement, according to the human metabolome database (HMDB, http://www.hmdb.ca/).
Statistical analysis: Peak extraction, alignment, normalization, and standardization were performed using a "home-built" code by MATLAB (R2016a, The Mathworks, Natick, MA).
Other univariate statistical analyses in this work were performed using SPSS software version 19.0 (IBM Corp., Armonk, New York), including one-sided DeLong test for AUC comparison, two-sided Student's t-test for age comparison, and Chi-square test for sex comparison. All significance level was set as 5%. The similarity score between two mass spectra was calculated by cosine correlation method in python following a reported algorithm. [10] To ensure to reach a statistical power > 0.9, power analysis was performed based on the pilot study by PASS            . The color scale ranging from -1 (blue) to 1 (red) indicated normalized mass spectrometric signal intensity in (a&b) and significant weight in (c), from low to high.     a) The pathway analysis was performed on MetaboAnalyst using the website built-in function; b) P value was calculated from the pathway topology analysis; c) Impact was calculated from pathway topology analysis.