Fitcsvm is slow when calling using engine.jar, but is quick when running directly through MATLAB R2018A

I am using engine.jar to be able to externally call MATLAB from another program (QuPath, for those wondering). I am using a function, with the parallel computing toolbox, which I also have enabled in the function. The function will complete in about 20 seconds when calling directly from MATLAB R2018A (and R2017A), but when I call via engine.jar, the function will take anywhere between 35-45 minutes to finish. Is there a limitation in which engine.jar would be throttled?

I am able to use the parallel computing function via the engine.jar, but it seems (for whatever reason) that there is some limitation in why the same data analyzed the same way would vary from seconds to about 45 minutes.

Code for reference:

% function classifier = train_classifier(rgb,idx,lumen,nuclei,stroma,cytoplasm)

% decision tree structure (first num vs. second num)
TREE = {0, [1 2 3]; ...     % lumen
      1, [2 3]; ...       % nuclei
      2, 3};              % stroma vs cytoplasm

% SVM parameters
KSVM = 1e4;                 % number of points per class for training            
FCN = 'polynomial';         % polynomial or rbf
ORDER = 1;                  % polynomial order
SIGMA = 1;                  % rbf sigma
BOXC = 0;                   % box constraint exponent (10^BOXC)
MAXITER = 10^6;             % maximum number of iterations

%%Training

% load cluster and assignment data
idx = idx(:);
assigned{1} = lumen;
assigned{2} = nuclei;
assigned{3} = stroma;
assigned{4} = cytoplasm;
clear lumen stroma nuclei cytoplasm centroid;

% transform HSV from cylindrical to cartesian coordinates
hsv = rgb2hsv(rgb);
[hsvc(:,:,1),hsvc(:,:,2),hsvc(:,:,3)] = pol2cart(2*pi*hsv(:,:,1),hsv(:,:,2),hsv(:,:,3));
clear hsv;

% change order
hsvc = permute(hsvc,[3 1 2]);
hsvc = hsvc([1 2 3],:);

% initialize
ntree = length(TREE);
classifier = NaN(ntree,4);

% train classifier
for i = 1:ntree

    % establish training data
    grp1 = find(ismember(idx,[assigned{TREE{i,1}+1}]));
    if length(grp1)>KSVM
        grp1 = randsample(grp1,KSVM,false);
    end
    grp2 = find(ismember(idx,[assigned{TREE{i,2}+1}]));
    if length(grp2)>KSVM
        grp2 = randsample(grp2,KSVM,false);
    end
    labels = [ones(length(grp1),1); zeros(length(grp2),1)];

    % perform SVM
    warning off;
    svm = fitcsvm([hsvc(:,grp1)'; hsvc(:,grp2)'], ...
                   labels, ...
                   'KernelFunction', FCN, ...
                   'PolynomialOrder', ORDER, ...
                   'BoxConstraint', 10^BOXC, ...
                   'IterationLimit', MAXITER, ...
                   'KernelScale','auto', ...
                   'HyperparameterOptimizationOptions',struct('UseParallel',true));
    warning on;
    clear grp1 grp2 labels;

    % update classifier
    classifier(i,:) = svmhyperplane(svm);
    clear svm;

end