function [prob,died]=physionet2012(tm,category,val)
% 
% Submission #1 for the PhysioNet 2012 Challenge, Phase 1.
% 
% Written by Alexandros Pantelopoulos, 23/04/2012
%
%
% Inputs:
% -------
%  tm      - (Nx1 Cell Array) Cell array containing time of measurement
%  category- (Nx1 Cell Array) Cell array containing type (category)
%            measurement
%  value   - (Nx1 Cell Array) Cell array containing value of measurement
%
%
% Outputs:
% --------
%  prob      - (Scalar) Probability value of the patient dying in the
%              hospital.
%  died      - (Logical) Binary classification if the patient is going to
%              die in the hospital (1 - Died, 0 - Survived)
% 
%
% Requirements:
% -------------
%  mat file 'nnet.mat'
%
%
% Description:
% ------------
%  My submission calculates the in-hospital death probability using several
%  simple statistical features extracted from the timeseries data and a
%  neural network classifier. In order to run it requires the mat file
%  'nnet.mat' to be in the same folder. This mat file contains the neural
%  network object created by MATLAB after training the classifier. It calls
%  the basic MATLAB function 'sim' in order to calculate the probability of
%  death in the hospital. In the case this probability is bigger than 0.5,
%  then the prediction is made that the patient will die in the hospital.
%  In this version the timestamps of the measurements are not utilized in
%  the classifier. Finally this function normalizes all features based on
%  the mean and stdev values calculated from the training set.
%

% Load the neural network object & the saved mean & stdev values
load('nnet.mat')

% Basic attribute descriptions
attributes = {'RecordID','Age','Gender','Height','Weight','Albumin','ALP','ALT','AST','Bilirubin','BUN','Cholesterol',...
    'Creatinine','DiasABP','FiO2','GCS','Glucose','HCO3','HCT','HR','K','Lactate','Mg','MAP','MechVent','Na','NIDiasABP',...
    'NIMAP','NISysABP','PaCO2','PaO2','pH','Platelets','RespRate','SaO2','SysABP','Temp','TroponinI','TroponinT','Urine','WBC'};
numAtributes = length(attributes);

% feature array initialization
numStaticF = 4;
numDynamicF = 36;
numStatsPerA = 5;
numFeatures = numDynamicF*numStatsPerA + numStaticF;
features = zeros(numFeatures,1);

% Extract all the attributes in individual cells
for j = 1:numAtributes
	eval([attributes{j} ' = val(strcmp(category,attributes(j)));'])
end

% Extract all the features from the existing attributes
ctr = 0;
% Age
j = 2;
eval(['temp = ' attributes{j} ';']);
if (length(temp)>1)
    ctr =  ctr + 1;
    features(ctr) = max(temp);
else
    ctr =  ctr + 1;
    features(ctr) = temp;
    if (temp == -1) % missing value
        features(ctr) = MEAN(ctr);
    end
end
% Gender and Height
for j = 3:4
    ctr = ctr + 1;
    eval(['temp = ' attributes{j} ';']);
    features(ctr) = temp;
    if (temp == -1) % missing value
        features(ctr) = MEAN(ctr);
    end
end
% MechVent
j = 25;
eval(['temp = ' attributes{j} ';']);
ctr =  ctr + 1;
if (isempty(temp))
    features(ctr) = 0;
else
    features(ctr) = max(temp);
end
% Grab the rest of he features from the remaining attributes
for j = 5:numAtributes
    if (strcmp(attributes{j},'MechVent') == 0)
        temp = 0;
        eval(['temp = ' attributes{j} ';']);
        if (isempty(temp))||(sum(temp == -1)>0) % missing value
            ctr = ctr + 1;
            features(ctr) = MEAN(ctr);
            ctr = ctr + 1;
            features(ctr) = MEAN(ctr);
            ctr = ctr + 1;
            features(ctr) = MEAN(ctr);
            ctr = ctr + 1;
            features(ctr) = MEAN(ctr);
            ctr = ctr + 1;
            features(ctr) = MEAN(ctr);
        else
            ctr = ctr + 1;
            eval(['features(ctr) = mean(' attributes{j} ');']);
            ctr = ctr + 1;
            eval(['features(ctr) = min(' attributes{j} ');']);
            ctr = ctr + 1;
            eval(['features(ctr) = max(' attributes{j} ');']);
            ctr = ctr + 1;
            eval(['features(ctr) = sqrt(var(' attributes{j} '));']);
            ctr = ctr + 1;
            eval(['features(ctr) = sum(diff(' attributes{j} '));']);
        end
    end
end

% Normalize the features
features = (features-MEAN)./STDEV;

% Calculate the probability of in-hospital death
prob = sim(nnet,features);

% Calculate the binary prediction of in-hospital death
TH = 0.5;
died = prob>=TH;

% Scale the prob value (emprically calculated based on the training set
prob = prob.^2;