OSCR

Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers.

Code ↔ Paper

2 matches between paragraphs of the paper and lines of its authors' code, computed by the harvester (lexical-v1). Click a colored paragraph or line to see its counterpart.

The 2 matches
  1. [1] § Methods › Ridge regression CPM ↔ train_model_ranked_edges.m, lines 221–259 · score 0.55 · hyperparameter optimization, fitrlinear, fitting, training, ridge, CPM
  2. [2] § Methods › Datasets ↔ run_models.m, lines 38–46 · score 0.51 · Card Sort age, Flanker, NIH, HBN

Paper

Loaded from Europe PMC by your browser, not stored by OSCR: doi.org · Europe PMC

The paper is loaded when this pane is shown.

The authors' code

MATLAB · 499 lines · 19 KB · CC-BY-4.0 · 1 match

  1. function results = train_model(dataset, varargin)
  2. %{
  3. Function to train a ridge CPM or basic CPM model
  5. To use this script for within dataset predictions, you should pass one
  6. dataset
  8. To use this script for external (cross-dataset) predictions, you should
  9. pass two datasets
  11. %}
  13. %% Parse input arguments
  14. p = inputParser;
  15. addRequired(p, 'dataset', @isstruct);
  16. addParameter(p, 'external_dataset', '', @isstruct); % external pca table
  17. addParameter(p, 'model_type', 'ridge', @ischar); % options: ridge, cpm
  18. addParameter(p, 'seed', 1, @isnumeric);
  19. addParameter(p, 'num_folds', 10, @isnumeric);
  20. addParameter(p, 'feat_thresh', 0.05, @isnumeric);
  21. addParameter(p, 'feat_selection', 'percent', @ischar); % percent, p, ranked_10_percent, ranked_20_percent, ranked_5_percent, ranked_1_percent, all, top_half, bottom_half
  22. addParameter(p, 'null', 0, @isnumeric); % 1 for yes
  23. addParameter(p, 'control_covars', 0, @isnumeric); % 1 for yes
  24. addParameter(p, 'ranked_segment', 1, @isnumeric);
  25. addParameter(p, 'lambda', NaN, @isnumeric);
  27. parse(p, dataset, varargin{:});
  29. external_dataset = p.Results.external_dataset;
  30. model_type = p.Results.model_type;
  31. seed = p.Results.seed;
  32. num_folds = p.Results.num_folds;
  33. feat_thresh = p.Results.feat_thresh;
  34. feat_selection = p.Results.feat_selection;
  35. null = p.Results.null;
  36. control_covars = p.Results.control_covars == 1;
  37. ranked_segment = p.Results.ranked_segment;
  38. lambda = p.Results.lambda;
  40. if ~isnan(lambda) && length(lambda) > 1
  41. error(['Only a single lambda value is supported. ', ...
  42. 'To optimize lambda automatically, omit the lambda parameter or set it to NaN.']);
  43. end
  46. clearvars p
  48. %% get data
  50. % matrix data (vectorized)
  51. X = mat2edge(dataset.mats)';
  53. % behavioral data
  54. num_behav = length(dataset.phenotypes);
  55. behav_all = dataset.behav_table{:, dataset.phenotypes};
  56. n = length(behav_all);
  58. % covariate data
  59. if control_covars
  60. if isfield(dataset, 'covars') && isfield(dataset, 'covar_table')
  61. covars = dataset.covar_table{:, dataset.covars};
  62. disp('running with covariates')
  63. else
  64. disp('control_covars is set to run covariates, but either covars phenotypes or covar_table not specified')
  65. end
  66. end
  68. % shuffle for null distribution
  69. rng(seed)
  70. if null == 1
  71. shuffle_idx = randperm(length(behav_all));
  72. behav_all = behav_all(shuffle_idx, :);
  73. if control_covars && isfield(dataset, 'covars') && isfield(dataset, 'covar_table')
  74. covars = covars(shuffle_idx, :);
  75. end
  76. end
  78. % option for no PCA, external PCA, or cross-validated PCA
  79. if length(dataset.phenotypes) == 1
  80. pca_type = 'none';
  81. behav = behav_all;
  82. elseif (length(dataset.phenotypes) > 1) && (isfield(dataset, 'external_behav_table'))
  83. % if more than 1 phenotype present and if external pca file exists
  84. pca_type = 'external';
  86. % run pca in external (non-imaging) data
  87. behav_external_pca = dataset.external_behav_table{:, dataset.phenotypes};
  88. [pca_coeff, score, latent, tsquared, explained, mu] = pca(behav_external_pca);
  90. % project behavior in main dataset
  91. behav_pcs = (behav_all - mu) * pca_coeff;
  92. behav = behav_pcs(:, 1);
  94. elseif (length(dataset.phenotypes) > 1) && (~isfield(dataset, 'external_behav_table'))
  95. % if more than 1 phenotype present but no external pca file
  96. if ~isstruct(external_dataset)
  97. % do PCA within cross-validation folds for within-dataset predictions
  98. pca_type = 'cv';
  100. % make an empty array to fill in with PC test data (obtained via PCA within cross-validation scheme)
  101. behav = NaN + zeros(n, 1);
  103. elseif isstruct(external_dataset)
  104. % if no behavior-only pca file is provided, you can still do internal PCA when predicting in an external dataset
  105. pca_type = 'internal';
  107. % run pca in external (non-imaging) data
  108. [pca_coeff, score, latent, tsquared, explained, mu] = pca(behav_all);
  110. % project behavior in main dataset
  111. behav_pcs = (behav_all - mu) * pca_coeff;
  112. behav = behav_pcs(:, 1);
  113. end
  114. end
  116. rng(seed)
  118. %% Within-dataset
  119. if isempty(external_dataset)
  121. % Select cross-validation splits
  122. cv_idx = cv_indices(n, num_folds);
  123. coef_all = zeros(size(X, 2), num_folds);
  124. coef0_all = zeros(num_folds, 1);
  125. y_predict = zeros(n, 1);
  126. lambda_total = NaN(num_folds, 1);
  128. for k = 1:num_folds
  130. train_idx = find(cv_idx ~= k);
  131. test_idx = find(cv_idx == k);
  133. X_train = X(train_idx, :);
  134. X_test = X(test_idx, :);
  136. if strcmp(pca_type, 'cv')
  137. % run pca in training data
  138. [pca_coeff, score, latent, tsquared, explained, mu] = pca(behav_all(train_idx, :));
  140. % project behavior in training data
  141. behav_pcs = (behav_all(train_idx, :) - mu) * pca_coeff;
  142. behav_train = behav_pcs(:, 1);
  144. % project behavior in test data
  145. behav_pcs = (behav_all(test_idx, :) - mu) * pca_coeff;
  146. behav_test = behav_pcs(:, 1);
  147. behav(test_idx) = behav_test; % save the test cross-validated PC for later
  148. else
  149. behav_train = behav(train_idx);
  150. behav_test = behav(test_idx);
  151. end
  153. % Step 1 of feature selection: correlation
  154. if control_covars
  155. covars_train = covars(train_idx);
  156. % select features with partial correlation
  157. [edge_corr, edge_p] = partialcorr(X_train, behav_train, covars_train);
  158. else
  159. % select features with standard correlation
  160. [edge_corr, edge_p] = corr(X_train, behav_train);
  161. end
  163. % Step 2 of feature selection: thresholding
  164. if strcmp(feat_selection, 'p') % selecting by p value
  165. p_thresh = feat_thresh;
  166. feat_loc = find(edge_p < p_thresh);
  167. elseif strcmp(feat_selection, 'percent')
  168. p_thresh = prctile(edge_p, 100 * feat_thresh);
  169. feat_loc = find(edge_p < p_thresh);
  170. elseif strcmp(feat_selection, 'ranked_10_percent')
  171. total_features = length(edge_p);
  172. segment_size = round(total_features * 0.10);
  173. start_index = (ranked_segment - 1) * segment_size + 1;
  174. end_index = min(ranked_segment * segment_size, total_features);
  175. [~, sorted_indices] = sort(edge_p);
  176. feat_loc = sorted_indices(start_index:end_index);
  177. elseif strcmp(feat_selection, 'ranked_20_percent')
  178. total_features = length(edge_p);
  179. segment_size = round(total_features * 0.20);
  180. start_index = (ranked_segment - 1) * segment_size + 1;
  181. end_index = min(ranked_segment * segment_size, total_features);
  182. [~, sorted_indices] = sort(edge_p);
  183. feat_loc = sorted_indices(start_index:end_index);
  184. elseif strcmp(feat_selection, 'ranked_5_percent')
  185. total_features = length(edge_p);
  186. segment_size = round(total_features * 0.05);
  187. start_index = (ranked_segment - 1) * segment_size + 1;
  188. end_index = min(ranked_segment * segment_size, total_features);
  189. [~, sorted_indices] = sort(edge_p);
  190. feat_loc = sorted_indices(start_index:end_index);
  191. elseif strcmp(feat_selection, 'ranked_1_percent')
  192. total_features = length(edge_p);
  193. segment_size = round(total_features * 0.01);
  194. start_index = (ranked_segment - 1) * segment_size + 1;
  195. end_index = min(ranked_segment * segment_size, total_features);
  196. [~, sorted_indices] = sort(edge_p);
  197. feat_loc = sorted_indices(start_index:end_index);
  198. elseif strcmp(feat_selection, 'all')
  199. feat_loc = 1:length(edge_p);
  200. elseif strcmp(feat_selection, 'top_half')
  201. [~, sorted_indices] = sort(edge_p);
  202. feat_loc = sorted_indices(1:round(length(edge_p)/2));
  203. elseif strcmp(feat_selection, 'bottom_half')
  204. [~, sorted_indices] = sort(edge_p, 'descend');
  205. feat_loc = sorted_indices(1:round(length(edge_p)/2));
  206. elseif strcmp(feat_selection, 'top_10_percent')
  207. p_thresh = prctile(edge_p, 10);
  208. feat_loc = find(edge_p < p_thresh);
  209. elseif strcmp(feat_selection, 'bottom_90_percent')
  210. p_thresh = prctile(edge_p, 10);
  211. feat_loc = find(edge_p > p_thresh);
  212. elseif strcmp(feat_selection, 'top_20_percent')
  213. p_thresh = prctile(edge_p, 20);
  214. feat_loc = find(edge_p < p_thresh);
  215. elseif strcmp(feat_selection, 'bottom_80_percent')
  216. p_thresh = prctile(edge_p, 20);
  217. feat_loc = find(edge_p > p_thresh);
  218. end
  221. % model fitting
  223. if strcmp(model_type, 'ridge')
  224. if isnan(lambda)
  225. mdl = fitrlinear(X_train(:, feat_loc), behav_train, ...
  226. 'Learner', 'leastsquares', ...
  227. 'Regularization', 'ridge', ...
  228. 'OptimizeHyperparameters', {'Lambda'}, ...
  229. 'HyperparameterOptimizationOptions', struct('ShowPlots', false, 'Verbose', 0));
  230. coef = mdl.Beta;
  231. coef0 = mdl.Bias;
  232. lambda_total(k) = mdl.ModelParameters.Lambda;
  233. fprintf('Lambda fold %d: %.4f\n', k, lambda_total(k));
  234. else
  235. mdl = fitrlinear(X_train(:, feat_loc), behav_train, ...
  236. 'Learner', 'leastsquares', ...
  237. 'Regularization', 'ridge', ...
  238. 'Lambda', lambda);
  239. coef = mdl.Beta;
  240. coef0 = mdl.Bias;
  241. lambda_total(k) = mdl.Lambda;
  243. fprintf('Lambda fold %d: %.4f\n', k, lambda_total(k));
  244. end
  246. coef_all(feat_loc, k) = coef;
  247. coef0_all(k) = coef0;
  249. % prediction
  250. y_predict(test_idx) = X_test(:, feat_loc) * coef + coef0;
  252. elseif strcmp(model_type, 'cpm')
  253. % get positive and negative networks, and summarize into a single feature
  254. pos_network = feat_loc(edge_corr(feat_loc) > 0);
  255. neg_network = feat_loc(edge_corr(feat_loc) < 0);
  256. X_train_summary = sum(X_train(:, pos_network), 2) - sum(X_train(:, neg_network), 2);
  258. % fit model
  259. mdl = robustfit(X_train_summary, behav_train);
  261. % predict
  262. X_test_summary = sum(X_test(:, pos_network), 2) - sum(X_test(:, neg_network), 2);
  263. y_predict(test_idx) = mdl(1) + mdl(2) * X_test_summary;
  265. % save coefficients
  266. coef = zeros(length(edge_corr), 1);
  267. coef(pos_network) = mdl(2);
  268. coef(neg_network) = -mdl(2);
  269. coef0 = mdl(1);
  270. coef_all(:, k) = coef;
  271. coef0_all(k) = coef0;
  272. end
  273. end
  275. results.y_true = behav; % I changed from y to y_true
  276. results.y_predict = y_predict;
  277. results.cv_idx = cv_idx;
  279. coef_consensus_mean = zeros(35778, 1); % Initialize coef_consensus_mean with zeros
  282. for i = 1:size(coef_all, 1)
  283. if all(coef_all(i, :) ~= 0)
  284. coef_consensus_mean(i) = mean(coef_all(i, :));
  285. end
  286. end
  288. results.lambda = lambda_total;
  289. results.coef_all = coef_all;
  290. results.coef_consensus_mean = coef_consensus_mean;
  291. results.coef_mean = mean(coef_all, 2);
  292. results.coef0_mean = mean(coef0_all, 2);
  293. results.r = corr(results.y_true, results.y_predict);
  294. results.pca_type = pca_type;
  295. results.phenotypes = dataset.phenotypes;
  296. results.model_type = model_type;
  297. results.seed = seed;
  298. results.num_folds = num_folds;
  299. results.feat_thresh = feat_thresh;
  300. results.feat_selection = feat_selection;
  301. results.null = null; % 1 is for null
  302. results.control_covars = control_covars; % 1 is for yes
  303. if ismember(feat_selection, {'ranked_10_percent', 'ranked_20_percent', 'ranked_5_percent', 'ranked_1_percent'})
  304. results.ranked_segment = ranked_segment;
  305. else
  306. results.ranked_segment = 'none';
  307. end
  308. if control_covars
  309. results.covars = dataset.covars; % which variables were used as covariates
  310. else
  311. results.covars = 'none';
  312. end
  314. %% Across datasets
  315. else
  316. % get second (external) dataset X data
  317. X_external = mat2edge(external_dataset.mats)';
  318. behav_all_external = external_dataset.behav_table{:, external_dataset.phenotypes};
  320. % for second (external) y data, depends if PCA is needed
  321. if length(external_dataset.phenotypes)==1
  322. behav_external = behav_all_external;
  323. elseif (length(external_dataset.phenotypes)>1) && (isfield(external_dataset, 'external_behav_table'))
  324. % if more than 1 phenotype present and if external pca file exists
  325. pca_type = 'external';
  327. % run pca in external (non-imaging) data
  328. behav_external_external_pca = external_dataset.external_behav_table{:, external_dataset.phenotypes}; % added "external_" before both "datasets:
  329. [pca_coeff,score,latent,tsquared,explained,mu] = pca(behav_external_external_pca);
  331. % project behavior in main dataset
  332. behav_pcs_external = (behav_all_external-mu)*pca_coeff;
  333. behav_external = behav_pcs_external(:, 1);
  335. elseif (length(external_dataset.phenotypes)>1) && (~isfield(external_dataset, 'external_behav_table'))
  336. pca_type = 'internal';
  338. % run pca in external dataset
  339. behav_external_internal_pca = external_dataset.behav_table{:, external_dataset.phenotypes};
  340. [pca_coeff,score,latent,tsquared,explained,mu] = pca(behav_external_internal_pca);
  342. % project behavior in main dataset
  343. behav_pcs_external = (behav_external_internal_pca-mu)*pca_coeff;
  344. behav_external = behav_pcs_external(:, 1);
  345. end
  347. % Step 1 of feature selection: correlation
  348. if control_covars
  349. % select features with partial correlation
  350. [edge_corr, edge_p] = partialcorr(X, behav, covars);
  351. else
  352. % select features with standard correlation
  353. [edge_corr, edge_p] = corr(X, behav);
  354. end
  356. % Step 2 of feature selection: thresholding
  357. if strcmp(feat_selection, 'p')
  358. p_thresh = feat_thresh;
  359. feat_loc = find(edge_p < p_thresh);
  360. elseif strcmp(feat_selection, 'percent')
  361. p_thresh = prctile(edge_p, 100 * feat_thresh);
  362. feat_loc = find(edge_p < p_thresh);
  363. elseif strcmp(feat_selection, 'ranked_10_percent')
  364. total_features = length(edge_p);
  365. segment_size = round(total_features * 0.10);
  366. start_index = (ranked_segment - 1) * segment_size + 1;
  367. end_index = min(ranked_segment * segment_size, total_features);
  368. [~, sorted_indices] = sort(edge_p);
  369. feat_loc = sorted_indices(start_index:end_index);
  370. elseif strcmp(feat_selection, 'ranked_20_percent')
  371. total_features = length(edge_p);
  372. segment_size = round(total_features * 0.20);
  373. start_index = (ranked_segment - 1) * segment_size + 1;
  374. end_index = min(ranked_segment * segment_size, total_features);
  375. [~, sorted_indices] = sort(edge_p);
  376. feat_loc = sorted_indices(start_index:end_index);
  377. elseif strcmp(feat_selection, 'ranked_5_percent')
  378. total_features = length(edge_p);
  379. segment_size = round(total_features * 0.05);
  380. start_index = (ranked_segment - 1) * segment_size + 1;
  381. end_index = min(ranked_segment * segment_size, total_features);
  382. [~, sorted_indices] = sort(edge_p);
  383. feat_loc = sorted_indices(start_index:end_index);
  384. elseif strcmp(feat_selection, 'ranked_1_percent')
  385. total_features = length(edge_p);
  386. segment_size = round(total_features * 0.01);
  387. start_index = (ranked_segment - 1) * segment_size + 1;
  388. end_index = min(ranked_segment * segment_size, total_features);
  389. [~, sorted_indices] = sort(edge_p);
  390. feat_loc = sorted_indices(start_index:end_index);
  391. elseif strcmp(feat_selection, 'all')
  392. feat_loc = 1:length(edge_p);
  393. elseif strcmp(feat_selection, 'top_half')
  394. [~, sorted_indices] = sort(edge_p);
  395. feat_loc = sorted_indices(1:round(length(edge_p)/2));
  396. elseif strcmp(feat_selection, 'bottom_half')
  397. [~, sorted_indices] = sort(edge_p, 'descend');
  398. feat_loc = sorted_indices(1:round(length(edge_p)/2));
  399. elseif strcmp(feat_selection, 'top_10_percent')
  400. p_thresh = prctile(edge_p, 10);
  401. feat_loc = find(edge_p < p_thresh);
  402. elseif strcmp(feat_selection, 'bottom_90_percent')
  403. p_thresh = prctile(edge_p, 10);
  404. feat_loc = find(edge_p > p_thresh);
  405. elseif strcmp(feat_selection, 'top_20_percent')
  406. p_thresh = prctile(edge_p, 20);
  407. feat_loc = find(edge_p < p_thresh);
  408. elseif strcmp(feat_selection, 'bottom_80_percent')
  409. p_thresh = prctile(edge_p, 20);
  410. feat_loc = find(edge_p > p_thresh);
  411. end
  414. % model training
  415. if strcmp(model_type, 'ridge')
  416. if isnan(lambda)
  417. mdl = fitrlinear(X_train(:, feat_loc), behav_train, ...
  418. 'Learner', 'leastsquares', ...
  419. 'Regularization', 'ridge', ...
  420. 'OptimizeHyperparameters', {'Lambda'}, ...
  421. 'HyperparameterOptimizationOptions', struct('ShowPlots', false, 'Verbose', 0));
  422. coef = mdl.Beta;
  423. coef0 = mdl.Bias;
  424. lambda_used = mdl.ModelParameters.Lambda;
  427. fprintf('Lambda used (optimized): %.4f\n', lambda_used);
  428. else
  429. mdl = fitrlinear(X(:, feat_loc), behav, ...
  430. 'Learner', 'leastsquares', ...
  431. 'Regularization', 'ridge', ...
  432. 'Lambda', lambda);
  433. coef = mdl.Beta;
  434. coef0 = mdl.Bias;
  435. lambda_used = mdl.Lambda;
  436. fprintf('Lambda used (fixed): %.4f\n', lambda_used);
  437. end
  440. elseif strcmp(model_type, 'cpm')
  441. pos_network = feat_loc(edge_corr(feat_loc) > 0);
  442. neg_network = feat_loc(edge_corr(feat_loc) < 0);
  443. X_summary = sum(X(:, pos_network), 2) - sum(X(:, neg_network), 2);
  445. % fit model
  446. mdl = robustfit(X_summary, behav);
  448. % save coefficients
  449. coef = zeros(length(edge_corr), 1);
  450. coef(pos_network) = mdl(2);
  451. coef(neg_network) = -mdl(2);
  452. coef0 = mdl(1);
  453. end
  455. % prediction
  456. if strcmp(model_type, 'ridge')
  457. y_predict = X_external(:, feat_loc) * coef + coef0;
  459. elseif strcmp(model_type, 'cpm')
  460. y_predict = X_external * coef + coef0;
  461. end
  463. % store results
  464. results.y_train = behav;
  465. results.y_external_true = behav_external;
  466. results.y_external_predict = y_predict;
  467. results.coef = coef;
  468. results.coef0 = coef0;
  471. if strcmp(model_type, 'ridge')
  472. results.lambda = lambda_used;
  473. elseif strcmp(model_type, 'cpm')
  474. results.lambda = NaN;
  475. end
  477. results.coef = coef;
  478. results.coef0 = coef0;
  479. results.r = corr(results.y_external_true, results.y_external_predict);
  480. results.pca_type = pca_type;
  481. results.phenotypes = dataset.phenotypes;
  482. results.model_type = model_type;
  483. results.seed = seed;
  484. results.feat_thresh = feat_thresh;
  485. results.feat_selection = feat_selection;
  486. results.null = null; % 1 is for null
  487. results.control_covars = control_covars; % 1 is for yes
  488. if ismember(feat_selection, {'ranked_10_percent', 'ranked_20_percent', 'ranked_5_percent', 'ranked_1_percent'})
  489. results.ranked_segment = ranked_segment; % save the rankem segment number if used
  490. else
  491. results.ranked_segment = 'none';
  492. end
  493. if control_covars
  494. results.covars = dataset.covars; % which variables were used as covariates
  495. else
  496. results.covars = 'none';
  497. end
  498. end
  499. end

train_model_ranked_edges.m at commit 9b10607, under CC-BY-4.0 · at the source

Overview

Authors: Brendan D. Adkinson1, Matthew Rosenblatt2, Huili Sun1, Javid Dadashkarimi3, Link Tejavibulya1, Corey Horien4, Margaret L. Westwater5,6, Raimundo X. Rodriguez1, Stephanie Noble7,8,9, Dustin Scheinost1,2,5,10,11,12
ORCID iDs: Brendan D. Adkinson, Matthew Rosenblatt, Huili Sun, Link Tejavibulya, Corey Horien, Raimundo X. Rodriguez, Stephanie Noble, Dustin Scheinost
  1. Yale School of Medicine,New Haven, CT USA
  2. Department of Biomedical Engineering, Yale University,New Haven, CT USA
  3. Department of Radiology, Perelman School of Medicine,Philadelphia, PA USA
  4. Department of Psychiatry, University of Pennsylvania,Philadelphia, PA USA
  5. Department of Radiology & Biomedical Imaging, Yale School of Medicine,New Haven, CT USA
  6. Department of Psychiatry, University of Oxford, Warneford Hospital,Oxford, UK
  7. Department of Bioengineering, Northeastern University,Boston, MA USA
  8. Department of Psychology, Northeastern University,Boston, MA USA
  9. Institute for Cognitive & Behavioral Health, Northeastern University,Boston, MA USA
  10. Department of Statistics & Data Science, Yale University,New Haven, CT USA
  11. Child Study Center, Yale School of Medicine,New Haven, CT USA
  12. Wu Tsai Institute, Yale University,New Haven, CT USA
Institutions: Yale University (United States); University of Pennsylvania (United States); Warneford Hospital (United Kingdom); Northeastern University (United States)
Journal: Nature human behaviour, volume 10, issue 7, pages 1356-1370
Dates: received 8 April 2025; accepted 16 March 2026; published online 15 April 2026; in print 2026
Type: Research article · Language: English
License: CC BY
Identifiers: DOI 10.1038/s41562-026-02447-y · PMID 41986741 · PMCID PMC13388108 · OpenAlex W7154479054
Open access: hybrid, a free copy (OpenAlex)
Status: code verified
Categories: structural MRI / diffusion (modality), human (organism), clinical / translational (subfield)
Methods: Smoothing, state filtering, decompositions, Machine learning, Statistics, Connectivity, fMRI & imaging
Keywords: Cognitive neuroscience, Machine learning, Psychology, Computational neuroscience
MeSH: Brain*, Connectome*, Machine Learning*, Neuroimaging*, Biomarkers, Diffusion Tensor Imaging, Female, Humans, Magnetic Resonance Imaging, Predictive Learning Models (* major topic)
Topic: Functional Brain Connectivity Studies (Cognitive Neuroscience, Neuroscience), according to OpenAlex
Funding: National Institute on Minority Health and Health Disparities (1F30MD018941); NIH (T32GM136651); NSF (DGE2139841); Gruber Foundation; National Institute of Mental Health (K00MH122372)
Citations: cited by 1 paper (Europe PMC); 124 references in the paper

Abstract

A central objective in human neuroimaging is to understand the neurobiology underlying cognition and mental health. Machine learning models trained on neuroimaging data are increasingly used as tools for predicting behavioural phenotypes, enhancing precision medicine and improving generalizability compared with traditional MRI studies. However, the high dimensionality of brain connectivity data makes model interpretation challenging. Prevailing practices rely on selecting features and, implicitly, interpreting identified feature networks as uniquely representative of a given phenotype while overlooking others. Despite its widespread use, how univariate feature selection balances the trade-off between simplification for optimizing modelling and oversimplification that misrepresents true neurobiology remains understudied. Here, using four large-scale neuroimaging datasets spanning over 12,000 participants and 13 outcomes, we demonstrate that edges discarded by feature selection can achieve significant prediction accuracies while yielding different neurobiological interpretations. These results are observed across cognitive, developmental and psychiatric phenotypes, extend to both functional connectivity (functional MRI) and structural (diffusion tensor imaging) connectomes, and remain evident in external validation. They suggest that focusing on only the top features may simplify the neurobiological bases of brain–behaviour associations. Such interpretations present only the tip of the iceberg when certain disregarded features may be just as meaningful, potentially contributing to ongoing issues surrounding reproducibility within the field. More broadly, our results reinforce that subtle brain-wide signals should not be ignored.

Reproduced under the paper's license (CC BY), from the paper cited above.

Repository

Its files are read in the Code ↔ Paper reader above, with 2 matches between paragraphs and lines of code.

brendan-adkinson/overlooked-features

License: CC-BY-4.0
State: the link answers, verified on 29 September 2026
Evidence: files inventoried
Commit: 9b1060783d708e71ee5b5714dd2a8d3a73baf139, 26 September 2025
Languages: MATLAB (4)
Size: 6 files, 4 scripts
Software Heritage: not archived
Found in: “Code availability”
Holds: README
Not found: license file, CITATION.cff, environment file, tests, continuous integration, documentation
Tools: Statistics and Machine Learning Toolbox (1 file)
Availability: 1 check, the latest on 29 September 2026: the link answers
  • 29 September 2026: the link answers
5 files
At the source: github.com/brendan-adkinson/overlooked-features

Code availability

The analyses were conducted using MATLAB v.R2024a. The code is available via GitHub at https://github.com/brendan-adkinson/overlooked-features. Preprocessing was carried out using Bioimage Suite v.3.01, which is freely available (https://medicine.yale.edu/bioimaging/suite/). Additional preprocessing was performed with the Human Connectome Project minimal preprocessing pipeline v.3.4.0 (https://github.com/Washington-University/HCPpipelines/releases).

Reproduced under the paper's license (CC BY), from the paper cited above.

Tracing map

Proposed by the machine: these links were found in the paper and verified at the source, without human review. The map will receive a Zenodo DOI once one of the paper's authors has validated it with their ORCID.

What the map holds:

  • 1 repository of the authors' code, each at its verified commit, with its license and how the link was found in the paper;
  • 4 scripts, each with its path and the digest of its content;
  • 2 matches between paragraphs of the paper and lines of the code (method lexical-v1);
  • neither the text of the paper nor the code itself.

Its JSON (tracing-map.json) is deposited on Zenodo with its DOI once the map is validated.

Data

No dataset and no data link were found in the paper.

Data availability

The data are available through the PNC, HCPD, HBN and ABCD datasets.

Reproduced under the paper's license (CC BY), from the paper cited above.

Versions

The history of this record: each version stored by the harvester or made by a correction of its authors or of the maintainers of its code, and what changed in its facts. The texts of the paper (its abstract, its availability statements) are not part of it; versions that changed only those are not listed.

Version 1, 29 September 2026: the first record

Recorded: type, language, journal, volume, issue, pages, dates, 10 authors, 4 keywords, 10 MeSH terms, 5 funders, 122 references.

Cite

This paper

Adkinson, B. D., Rosenblatt, M., Sun, H., Dadashkarimi, J., Tejavibulya, L., Horien, C., Westwater, M. L., Rodriguez, R. X., Noble, S., & Scheinost, D. (2026). Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers. Nature human behaviour, 10(7), 1356-1370. https://doi.org/10.1038/s41562-026-02447-y

BibTeX

@article{adkinson2026feature,
author = {Adkinson, Brendan D. and Rosenblatt, Matthew and Sun, Huili and Dadashkarimi, Javid and Tejavibulya, Link and Horien, Corey and Westwater, Margaret L. and Rodriguez, Raimundo X. and Noble, Stephanie and Scheinost, Dustin},
title = {{Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers}},
journal = {Nature human behaviour},
year = {2026},
month = apr,
volume = {10},
number = {7},
pages = {1356--1370},
publisher = {Nature Portfolio},
issn = {2397-3374},
doi = {10.1038/s41562-026-02447-y},
url = {https://doi.org/10.1038/s41562-026-02447-y},
pmid = {41986741},
pmcid = {PMC13388108}
}

RIS

TY - JOUR
AU - Adkinson, Brendan D.
AU - Rosenblatt, Matthew
AU - Sun, Huili
AU - Dadashkarimi, Javid
AU - Tejavibulya, Link
AU - Horien, Corey
AU - Westwater, Margaret L.
AU - Rodriguez, Raimundo X.
AU - Noble, Stephanie
AU - Scheinost, Dustin
TI - Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers
T2 - Nature human behaviour
J2 - Nat Hum Behav
PY - 2026
DA - 2026/04/15
VL - 10
IS - 7
SP - 1356
EP - 1370
SN - 2397-3374
PB - Nature Portfolio
DO - 10.1038/s41562-026-02447-y
UR - https://doi.org/10.1038/s41562-026-02447-y
LA - en
ER -

CSL-JSON

{
"id": "10.1038/s41562-026-02447-y",
"type": "article-journal",
"title": "Feature selection leads to divergent neurobiological interpretations of brain-based machine learning biomarkers",
"container-title": "Nature human behaviour",
"author": [
{
"family": "Adkinson",
"given": "Brendan D."
},
{
"family": "Rosenblatt",
"given": "Matthew"
},
{
"family": "Sun",
"given": "Huili"
},
{
"family": "Dadashkarimi",
"given": "Javid"
},
{
"family": "Tejavibulya",
"given": "Link"
},
{
"family": "Horien",
"given": "Corey"
},
{
"family": "Westwater",
"given": "Margaret L."
},
{
"family": "Rodriguez",
"given": "Raimundo X."
},
{
"family": "Noble",
"given": "Stephanie"
},
{
"family": "Scheinost",
"given": "Dustin"
}
],
"container-title-short": "Nat Hum Behav",
"volume": "10",
"issue": "7",
"page": "1356-1370",
"DOI": "10.1038/s41562-026-02447-y",
"PMID": "41986741",
"PMCID": "PMC13388108",
"ISSN": "2397-3374",
"publisher": "Nature Portfolio",
"URL": "https://doi.org/10.1038/s41562-026-02447-y",
"language": "en",
"issued": {
"date-parts": [
[
2026,
4,
15
]
]
}
}

The tracing map gets a citation of its own once an author has validated it and it has a DOI.

Similar papers

The papers with a page that share the most with this one: the tools found in their code, their categories, datasets, cited references and authors, the rarest counting most.

[1] doi:10.1038/s44220-026-00623-7 [code]
Optimizing functional connectivity scanning conditions for predicting autistic traits.
Journal: Nature. Mental health
In common: Statistics and Machine Learning Toolbox, clinical / translational, 12 references, 3 authors
[2] doi:10.1038/s41467-026-73941-0 [code]
Using connectome-based predictive models to reveal the systems standardized tests and clinical symptoms are reflecting.
Journal: Nature communications
In common: 10 references, author Dustin Scheinost
[3] doi:10.1038/s41467-026-75661-x [code]
A neural signature of sleep deprivation in the human brain.
Journal: Nature communications
In common: 12 references
[4] doi:10.1002/hbm.70493 [code]
Testing for Network Specificity in Brain-Behavior Associations Using Ordinal Dominance Curves.
Journal: Human brain mapping
In common: 8 references
[5] doi:10.1038/s41467-026-73668-y [code]
Convergent and divergent brain-cognition development in early adolescence.
Journal: Nature communications
In common: Statistics and Machine Learning Toolbox, 7 references
[6] doi:10.7554/elife.104053 [code]
Brain cognition gaps reveal associations with dopamine and factors related to brain health through artificial intelligence prediction of functional connectome.
Journal: eLife
In common: 7 references
[7] doi:10.1162/imag.a.1287 [code]
Is the whole more than the sum of its parts? Considering global and local features of the connectome improves prediction of individuals and phenotypes.
Journal: Imaging neuroscience (Cambridge, Mass.)
In common: Statistics and Machine Learning Toolbox, 6 references
[8] doi:10.1038/s41593-026-02359-0 [code]
The cross-site reproducibility of MRI morphometric phenotypes in psychiatric disorders.
Journal: Nature neuroscience
In common: Statistics and Machine Learning Toolbox, structural MRI / diffusion, 6 references
[9] doi:10.1038/s41467-026-73072-6 [code]
Mapping the spatiotemporal continuum of structural connectivity development across the human connectome in youth.
Journal: Nature communications
In common: structural MRI / diffusion, 7 references
[10] doi:10.1523/eneuro.0370-25.2026 [code]
Pretraining for Large-Scale Functional Connectome Fingerprinting Supports Generalization and Transfer Learning in Functional Neuroimaging.
Journal: eNeuro
In common: 7 references

Contribute

The authors of this paper can claim it, correct its record and validate its tracing map, and the maintainers of its code (its owner, or a public member of its organization) correct what it says of their repository; anyone signed in can ask for its removal. Every request goes to OSCR's own machine, which answers it; your account page follows them.

Sign in with ORCID to claim this paper as one of its authors, correct its record or validate its tracing map: when the paper's metadata lists your ORCID iD, you are recognized at once. Maintainers of its code: sign in with GitHub, then claim the repository on your account page.

Request its removal

To ask OSCR to remove this record, the copies of its authors' scripts or its tracing map, use the removal request page: signed in, you say who you are, what to remove and why, then review and confirm the request. Published rules decide every request (how).

Discussion, reproductions, activity

Discussion: questions and error reports about this paper and its code, from signed-in readers and its authors. It opens with sign-in.

Reproductions: reports from readers who ran the authors' code: what they reproduced, with which environment, commit and data. It opens with sign-in.

Activity: what happens around this paper: new versions of its record, its map's validation, discussions and reproductions. It opens with sign-in.