function LDA_classifier(DYN_DATA,IHD,time_series_names,MD_DATA_D,MD_DATA_DD,MD_DATA,DESCRIPTORS,des_names,R,MD_VARI,MEAN_DATA_24,MEAN_DATA_48)
% %Y=(IHD-1)+IHD; % classes y ={-1,1}
% % X=[MEAN_DATA_24 MEAN_DATA_48];
%
% %X=[MEAN_DATA_24(params_to_use)];
%
% % normeeraus ja keskiarvon poisto
%
%
% X=[MEAN_DATA_24 MEAN_DATA_48 ];
%
% %X=[X1];
%
% %X=[X];
% AGE_a=DESCRIPTORS(:,strcmp(des_names,'Age'));
% Gender_a=DESCRIPTORS(:,strcmp(des_names,'Gender'));
% ICUTYPE_a=[DESCRIPTORS(:,strcmp(des_names,'ICUType'))];
%
% ICUTYPE_t=ICUTYPE_a;
%
% ICUTYPE_a(ICUTYPE_t==1)=2;
% ICUTYPE_a(ICUTYPE_t==2)=4;
% ICUTYPE_a(ICUTYPE_t==3)=1;
% ICUTYPE_a(ICUTYPE_t==4)=3;
%
% X=[AGE_a ICUTYPE_a X];
%
%
%
%
%
% %if des{:,strcmp(des_names,'Gender')}==0
%
% %X_new_rec=zeros(size(X));
%
% savefile='gender0.mat';
% savefile='gender0_60.mat';
% load(savefile,'N2','PHAT_all','best_th','A','S','Mu','V', 'CV', 'HP', 'LC','N','pc1','mu','sigma','B');
%
%
% % S_new=calc_S_new_data(X',A,V,Mu,N,CV);
% X_new_rec =( repmat(Mu,1,size(S,2)) + (A*S))';
% % mu1=Mu;
% sigma0 = sigma;
% sigma0(sigma0==0) = 1;
% z = bsxfun(@minus,X_new_rec', mu');
% z = bsxfun(@rdivide, z, sigma0');
%
% % x_in=(pinv(pc1)*z)';
% % X(Gender_a==0,:)=x_in(Gender_a==0,:);
% z=z';
% X(Gender_a==0,:)=z(Gender_a==0,:);
%
% savefile='gender1.mat';
% savefile='gender1_60.mat';
% load(savefile,'N2','PHAT_all','best_th','A','S','Mu','V', 'CV', 'HP', 'LC','N','pc1','mu','sigma','B');
% % mu2=Mu;
% % plot(mu1)
% % hold on
% % plot(mu2)
% % S_new=calc_S_new_data(X(Gender_a==1,:),A,V,Mu,N,CV);
% X_new_rec =( repmat(Mu,1,size(S,2)) + A*S)';
% sigma0 = sigma;
% sigma0(sigma0==0) = 1;
% z = bsxfun(@minus,X_new_rec', mu');
% z = bsxfun(@rdivide, z, sigma0');
% % x_in=(pinv(pc1)*z)';
% % X(Gender_a==1,:)=x_in(Gender_a==1,:);
% z=z';
% X(Gender_a==1,:)=z(Gender_a==1,:);
%
% X=[ MD_DATA double(R) MD_VARI X];
% %X=[ MD_DATA MD_VARI X];
%
% R=~isnan(X);
% s=sum(R,2)==size(X,2); % selects only values that do not contain NaNs
% dis=sum(R,2)~=size(X,2);
%
% X=X(s,:);
% IHD=IHD(s);
% s_selector=s;
% %B=log_reg(X,Y,R);
% %B=teach_log_reg(X,Y,R);
%
%
%
% alpha = 0:0.05:0.7;
% lambda = 0:0.05:0.7;
% scores=zeros(length(alpha)*length(lambda),7);
%
% idx=1;
%
% n=size(X,1);
% s=zeros(n,1);
% s(1:round(n/2))=1;
% s=boolean(s);
% dataTrain=X(s,:);
% clabelsTrain=IHD(s);
% labelsTrain=~IHD(s)+1;
%
%
% dataTest=X;
% labelsTest=~IHD+1;
%
% dataValid=X(~s,:);
% labelsValid=~IHD(~s)+1;
matrixData=DYN_DATA;
clabels=IHD;
weightsLDA = getLDAweights(matrixData,clabels);
featureVectors = featureExtractionLDA(matrixData,weightsLDA);
X=featureVectors;
R=~isnan(X);
s=sum(R,2)==size(X,2); % selects only values that do not contain NaNs
dis=sum(R,2)~=size(X,2);
X=X(s,:);
IHD=IHD(s);
s_selector=s;
%B=log_reg(X,Y,R);
%B=teach_log_reg(X,Y,R);
alpha = 0:0.05:0.7;
lambda = 0:0.05:0.7;
scores=zeros(length(alpha)*length(lambda),7);
scores2=zeros(length(alpha)*length(lambda),7);
idx=1;
n=size(X,1);
s=zeros(n,1);
s(1:round(n/2))=1;
s=boolean(s);
dataTrain=X(s,:);
clabelsTrain=IHD(s);
labelsTrain=~IHD(s)+1;
dataTest=X;
labelsTest=~IHD+1;
dataValid=X(~s,:);
labelsValid=~IHD(~s)+1;
%% jukkiksen starts
for alpha=0:0.05:0.5
for lambda=0:0.05:0.5
D = size(dataTrain,2); % dimension
maxIter = 500;
w_init = 0.1*ones(D+1,1); % initialize parameters
% optimoi malli, siis data annetaan normaalisti piirrevektorina:
[w_opt,fx,it] = minimize(w_init, 'objFuncLR', maxIter, dataTrain, clabelsTrain, alpha, lambda);
% laske ennustustarkkuus:
[ResTrain,labelsEstimTrain,probsTrain,allProbsTrain] = predict_LRoma(w_opt,dataTrain,labelsTrain);
[ResValid,labelsEstimValid,probsValid,allProbsValid] = predict_LRoma(w_opt,dataValid,labelsValid);
[ResTest,labelsEstimTest,probsTest,allProbsTest] = predict_LRoma(w_opt,dataTest,labelsTest);
%% jukkiksen ends
%[best_th,max_score1,BEST_DATA]=opt_th(IHD(s),allProbsTrain);
[best_th1,max_score11,th1,max_score2_11,BEST_DATA]=opt_th(IHD(~s),allProbsValid);
%[best_th,max_score12,th2,max_score2_12,BEST_DATA]=opt_th(~IHD(~s),allProbsValid);
[best_th2,max_score21,th2,max_score2_21,BEST_DATA]=opt_th(IHD(s),allProbsTrain);
%[best_th,max_score22,th2,max_score2_22,BEST_DATA]=opt_th(~IHD(s),allProbsTrain);
scores(idx,:)=[alpha lambda best_th1 max_score11 th1 max_score21 0];
scores2(idx,:)=[alpha lambda best_th2 max_score2_11 th2 max_score2_21 0];
idx=idx+1;
end
end
scores=scores(1:idx-1,:);
scores2=scores2(1:idx-1,:);
[m,i]=max(scores(:,4));
save('score1_LDA.mat','scores')
[m,i]=min(scores2(:,4));
save('score2_LDA.mat','scores2')
function [best_th,max_score1,best_th2,max_score2,BEST_DATA]=opt_th(IHD_,P)
num_params=size(time_series_names,1);
DATA=zeros(size(IHD_,1),3);
%B = mnrfit(X_,IHD_+1);
PHAT = P';
max_score1=0;
max_score2=0;
best_th=0;
best_th2=0;
for class_th=.1:.01:0.5
DATA(:,1)=str2double('0000');
DATA(:,2)=PHAT(:,1);
DATA(:,3)=PHAT(:,1)> class_th;
DATA(DATA(:,2)<0.01,2)=0.01;
DATA(DATA(:,2)>0.99,2)=0.99;
% if(~isempty(results))
% Calculate sensitivity (Se) and positive predictivity (PPV)
TP=sum(DATA(IHD_==1,3));
FN=sum(~DATA(IHD_==1,3));
FP=sum(DATA(IHD_==0,3));
Se=TP/(TP+FN);
PPV=TP/(TP+FP);
show=0; % if show is 1, the decile graph will be displayed by lemeshow()
H=lemeshow([IHD_ DATA(:,2)],show);
% Use the title of figure to display the results
% title(['H= ' num2str(H) ' Se= ' num2str(Se) ' PPV= ' num2str(PPV) '. ' num2str(class_th) ])
% The event 1 score is the smaller of Se and PPV.
score1 = min(Se, PPV);
if score1>max_score1
max_score1=score1;
best_th=class_th;
% max_score2=H;
BEST_DATA=DATA;
% display(['Unofficial Event 1 score: ' num2str(score1)]);
end
if H>max_score2
max_score2=H;
best_th2=class_th;
% BEST_DATA=DATA;
% display(['Unofficial Event 1 score: ' num2str(score1)]);
end
% end
end
end
end