Created public meta-analysis toolbox MKDA

Meta_NBC/@meta_dataset/meta_dataset.m

@@ -0,0 +1,179 @@
+classdef meta_dataset
+    % Define a meta-analysis dataset object for classification analyses
+    %
+    % Properties:
+    % ---------------------------------------------------------------
+    %     dat               Data; sparse logical
+    %     connames          cell array of contrast names
+    %     volInfo           volInfo structure for 3-D reconstruction; see iimg_get_data.m
+    %     classes           Contrast classes; integers
+    %     classnames        Names for classes; cell of strings
+    %     params            Parameter structure
+    %     terms             Terms to classify
+    %     include_cases     Logical vector of contrasts to include
+    %     include_vox       Logical vector of voxels to include (features)
+    %     train_pred_fcn    Function handle to create predicted classes from binary data
+    %     test__pred_fcn    Function handle to create predicted classes
+    %     loss_fcn          Function handle of loss (error) function to evaluate
+    %   
+    %
+    % Class constructor: Creating a meta_dataset object
+    % ---------------------------------------------------------------
+    % obj = meta_dataset(varargin)
+    % varargin can be any property, value pair
+    %
+    % Examples:
+    % data = meta_dataset;
+    % data = meta_dataset('dat', MC_Setup.unweighted_study_data);
+    %
+    % data = meta_dataset('MC_Setup', MC_Setup); % special to convert MC_Setup structure
+    %
+    % Full classification example for pain vs. touch (from Neurosynth):
+    % data = meta_dataset('MC_Setup', MC_Setup);
+    % data = setup_data(data, {'pain', 'touch'});
+    % nbc = train(meta_nbc, data);
+    % nbc = test(nbc, data.dat);
+    % nbc = get_error(nbc, data.classes);
+    % report(nbc);
+    %
+    % Full classification example 2:
+    %   ---------------------------------------------------------
+    % DB = Meta_Setup(DB, 10);
+    % MC_Setup = Meta_Activation_FWE('setup', DB); % Here, choose Valence / positive negative
+    %            - you must enter a contrast (any contrast) across the
+    %            levels you wish to classify
+    % data = meta_dataset('MC_Setup', MC_Setup);
+    % data.params.min_vox = 50; % do not exclude sparse maps, just those < 50 vox
+    % data = setup_data(data, {'positive', 'negative'});
+    % % Cross-validate and get cv error rate
+    % nbc = cv(nbc, data);
+    % nbc = report(nbc);
+
+
+    properties
+
+        dat
+        connames
+        volInfo
+        classes
+        classnames
+        params
+        terms
+        include_cases
+        include_vox
+        train_pred_fcn
+        test_pred_fcn
+        loss_fcn
+
+
+    end
+
+    methods
+
+        % Class constructor
+        function obj = meta_dataset(varargin)
+            %
+            % obj = meta_dataset(varargin)
+
+            % ---------------------------------
+            % Create empty fmri_data object, and return if no additional
+            % arguments
+            % ---------------------------------
+
+            obj.dat = sparse(logical(false));
+            obj.connames = {};
+            obj.volInfo = [];
+            obj.classes  = single(0);
+            obj.classnames = {};
+            obj.terms = {};
+            obj.include_cases = logical(false);
+            obj.include_vox = logical(false);
+
+            obj.train_pred_fcn = [];
+            obj.test_pred_fcn = [];
+            obj.loss_fcn = [];
+
+            obj.params.min_vox = 5000;  % Studies must report at least this many sig voxels (after kernel conv)
+            obj.params.term_freq_cutoff = 0.001; % Minimum frequency of word use for study to be counted as involving term
+            obj.params.min_prop_active = 0.03;  % Voxels must have at least this proportion of active studies
+
+            % Varying the threshold has a substantial effect. A
+            % cut-off around .001 seems to work well for different
+            % datasets.
+            obj.params.predict_pval_thresh = 0.05;
+
+            obj.params.k = 10;  % number of cross-validation folds
+            obj.params.s_param = 2; % smoothing param for Naive Bayes m-estimator; Caffo says 2
+
+            % The code below can be generic to any class definition
+            % It parses 'fieldname', value pairs of inputs
+            % and returns a warning if unexpected strings are found.
+
+            if nargin == 0
+                return
+            end
+
+            % all valid fieldnames
+            valid_names = fieldnames(obj);
+            valid_names = [valid_names; {'MC_Setup', 'MC_setup', 'mc_setup'}'];
+
+            for i = 1:length(varargin)
+                if ischar(varargin{i})
+
+                    % Look for a field (attribute) with the input name
+                    wh = strmatch(varargin{i}, valid_names, 'exact');
+
+                    % behaviors for valid fields
+                    if ~isempty(wh)
+
+                        % special methods for specific fields
+                        switch varargin{i}
+
+                            % Re-cast vars as correct data types
+                            case {'include_cases' 'include_vox'}
+                                obj.(varargin{i}) = logical(varargin{i+1});
+
+                            case 'dat'
+                                obj.(varargin{i}) = sparse(logical(varargin{i+1}));
+
+                            case {'MC_Setup', 'MC_setup', 'mc_setup'}
+                                M = varargin{i+1};
+                                M.unweighted_study_data = sparse(logical(M.unweighted_study_data));
+                                obj.dat = M.unweighted_study_data;
+                                M = rmfield(M, 'unweighted_study_data'); % save memory space
+                                obj.volInfo = M.volInfo;
+
+                                %This worked for Neurosynth...below is mod
+                                %for MC_Setup from Meta_Activ_FWE
+                                %obj.connames = M.connames;
+
+                                obj.connames = cell(size(obj.dat, 2), 1);
+                                if isfield(M, 'Xi')
+                                    for j = 1:size(M.Xi, 2)
+                                        obj.connames(logical(M.Xi(:, j))) = M.Xinms(j);
+                                    end
+                                end
+
+                            otherwise
+                                obj.(varargin{i}) = varargin{i + 1};
+                        end
+
+                        % eliminate strings to prevent warnings on char
+                        % inputs
+                        if ischar(varargin{i + 1})
+                            varargin{i + 1} = [];
+                        end
+
+                    else
+                        warning('inputargs:BadInput', 'Unknown field: %s', varargin{i});
+
+                    end
+                end % string input
+            end % process inputs
+
+        end % constructor function
+
+    end % methods
+
+end % classdef
+