cs.LGOct 19, 2026

ITU-DESC: a missingness-aware machine-learning dementia risk score in the UK Biobank cohort

Authors: Abdullah Kavaklı, Faik Boray Tek

Organizations: Department of Artificial Intelligence and Data Engineering, Istanbul Technical University, Istanbul, Turkey.

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Background: Existing dementia risk scores rely on fixed, pre-selected features and linear models, or ignore bias from informative missingness in cohort data. Here we introduce the Istanbul Technical University Dementia Risk Score (ITU-DESC), an explainable machine-learning score for incident all-cause dementia that combines broad feature selection across roughly 8,000 candidate features with explicit control of missingness bias. Results: We studied approximately 500,000 UK Biobank participants with a median follow-up of 16.6 years. We withheld missingness patterns from the model using unconditional random hot-deck imputation. We stabilized feature selection by repeating BorutaSHAP-based feature selection with ten random-imputation seeds and selecting predictors chosen in at least eight of the ten seeds, which yielded 38 predictors. We trained a CatBoost gradient-boosted decision-tree model and evaluated it on a holdout test set of 50,041 participants. For all-cause incident dementia, ITU-DESC achieved an average precision of 0.138 (95% confidence interval 0.123-0.156) and an area under the receiver operating characteristic curve (ROC AUC) of 0.853 (0.844-0.863), significantly outperforming other risk scores, including the strongest comparator, the UK Biobank Dementia Risk Prediction model (0.118 and 0.841). We identified dementia risk predictors absent from established scores, including leukocyte telomere length and liver enzymes. Conclusions: ITU-DESC is the first dementia risk score to combine broad feature selection with explicit control of missingness bias.

Figures & tables

Figure 1: Participant selection for the present study from the UK Biobank cohort. The upper boxes show the number of participants remaining after exclusions and the split of the eligible cohort into a 90% development set and a 10% holdout test set. The lower boxes report counts separately for the development (Dev.) and holdout test sides of the split. For each side, they give the number of participants without incident dementia during follow-up. They also report the numbers of incident dementia and Alzheimer’s disease cases within the 5-year, 10-year and full follow-up periods.
Figure 2: Feature-selection pipeline. A stratified 90/10 split holds 10% of participants out as the test set. Within the 90% development set, the boxed two-stage pipeline consists of a five-fold SHAP screen that removes features with zero importance, followed by BorutaSHAP wrapper selection. The pipeline is repeated in full for each of the ten random-imputation seeds. Both stages therefore run ten times, and each seed carries its own screened feature set into its own wrapper run. Features selected in at least 8 of the 10 seeds form the ITU-DESC feature set. The model is then fitted on the development set and evaluated once on the holdout test set, which the selection stages never see.
ModelAP (95% CI)Δ\Delta AP (95% CI)PP
ITU-DESC (our model)0.138 (0.123, 0.156)——
UKB-DRP You et al. (2022)0.118 (0.105, 0.133)0.021 (0.008, 0.033)<0.01<0.01
UKBDRS Anatürk et al. (2023)0.102 (0.092, 0.114)0.036 (0.023, 0.050)<0.001<0.001
UKBDRS (w/o APOE ε\varepsilon 4)0.084 (0.076, 0.093)0.055 (0.041, 0.070)<0.001<0.001
DRS Walters et al. (2016)0.064 (0.058, 0.069)0.075 (0.061, 0.090)<0.001<0.001
CAIDE Kivipelto et al. (2006)0.044 (0.040, 0.048)0.094 (0.081, 0.110)<0.001<0.001
Table 1: Discrimination of ITU-DESC and established dementia risk scores for all-incident dementia 0 0 footnotetext: All models were evaluated on the same holdout test set ( n=50,041n=50{,}041 ; 1,1381{,}138 incident dementia cases), under identical outcome definitions, using isotonic-calibrated predictions. The upper panel reports average precision and the lower panel reports ROC AUC for the same models. Differences are ITU-DESC minus the comparator and were calculated from unrounded estimates. ROC AUC estimates, confidence intervals, differences and two-sided PP values were derived using DeLong’s method. AP confidence intervals were obtained by paired subject-level bootstrap resampling (10,000 resamples; percentile method), whereas AP PP values were obtained using two-sided paired permutation tests (2,000 permutations). AP, average precision; CI, confidence interval; ROC AUC, area under the receiver operating characteristic curve.
Figure 3: Comparison of ITU-DESC with established dementia risk scores. All models were evaluated on the holdout test set using calibrated predictions. a , Receiver operating characteristic curves for all-incident all-cause dementia ( 1,1381{,}138 incident cases). b , Precision–recall curves for all-incident all-cause dementia. c , Receiver operating characteristic curves for all-incident Alzheimer’s disease ( 513513 incident cases). d , Precision–recall curves for all-incident Alzheimer’s disease. The dotted line marks chance discrimination in a and c , and the outcome prevalence in b and d , which equals the average precision of a non-informative model. Discrimination estimates and pairwise comparisons are reported in Table 1 for dementia and Additional file 1: Table S12 for Alzheimer’s disease.
Figure 4: Discrimination of ITU-DESC across the 5-year, 10-year and all-incident horizons. A single fitted model underlies all three horizons, evaluated on the holdout test set using calibrated predictions. a , Receiver operating characteristic curves for all-cause dementia. b , Precision–recall curves for all-cause dementia. c , Receiver operating characteristic curves for Alzheimer’s disease. d , Precision–recall curves for Alzheimer’s disease. The dotted line marks chance discrimination in a and c . Discrimination values are reported in Additional file 1: Table S13, and the case counts underlying each horizon in Fig. 1 .
Figure 5: Flexible calibration of ITU-DESC for the all-incident outcomes. Both panels use the holdout test set. a , All-cause dementia. b , Alzheimer’s disease. Observed risk, estimated by locally estimated scatterplot smoothing (LOESS), is plotted against predicted risk with pointwise 95% confidence bands, calculated as the smoothed estimate plus or minus 1.96 standard errors of the local observed proportion. Histograms along the bottom of each panel show the distribution of predicted risks. Calibration slope, mean predicted risk against observed incidence and Brier score are given in Additional file 1: Table S23.
Figure 6: SHAP summary plot for ITU-DESC for all-incident dementia. Each point represents one participant and is plotted according to the feature’s SHAP value, corresponding to its signed contribution to the model output on the log-odds scale. Points are colored by feature value. Features are ordered from top to bottom by mean absolute SHAP value. SHAP values were computed on the holdout test set ( n=50,041n=50{,}041 ), using the model refit on the full development set.
Figure 7: Cumulative discrimination of ITU-DESC as predictors are added in SHAP-rank order for all-incident dementia. The x-axis lists the top- kk features ranked by mean absolute SHAP value computed on the 90% development set. a , Development-set SHAP importance of each feature. b , ROC AUC when the model is refit on the development set using the top- kk features and evaluated on the holdout test set ( n=50,041n=50{,}041 ). c , Precision–recall AUC (average precision) for the same refits. Each point is a single holdout-test evaluation.
Appendix figures & tables6 assets

Supplementary material from the paper’s appendix.

Appendix

 Development ( n=450,369n=450{,}369 )Holdout test ( n=50,041n=50{,}041 )
OutcomeCases%Cases%
Dementia, 5-year6810.151750.150
Dementia, 10-year3,8020.8444210.841
Dementia, all-incident10,2422.2741,1382.274
Alzheimer’s disease, 5-year2310.051260.052
Alzheimer’s disease, 10-year1,5060.3341670.334
Appendix
Table S1: Outcome counts and prevalences in the development and holdout test sets.
 DevelopmentHoldout test 
Variablenn%nn%Diff.
Sex (stratified on)
Female245,01854.40427,22554.4050.002
Male205,35145.59622,81645.5950.002
Age at recruitment (acceptance criterion)
Mean, years450,36956.54150,04156.4950.0057
Appendix
Table S2: Balance of the 90% development and 10% holdout test sets. Percentages are column percentages within each set; Diff. is the absolute difference in percentage points. For age at recruitment, a continuous variable, the % cells give the mean age in years and Diff. gives the standardized mean difference (SMD). For predictor availability, nn and % are the participants with a recorded value. Assessment centers are given as UK Biobank Field 54 codes. Ethnicity categories derive from UK Biobank Field 21000. The split was stratified on the six outcomes (Table S1 ) and sex. Age, ethnicity and assessment center were prespecified acceptance criteria checked after the split, not stratification variables. Predictor availability was reported but neither stratified on nor subject to a threshold.
StageOperationColumns
Raw extract—32,612
CompressionBaseline instance only; per-field collapse and typing11,926
Outcome removalDementia- and outcome-linked fields ( −586-586 )11,340
Expansion148 multiple-choice fields replaced by 26,501 indicators37,693
PreprocessingA priori exclusions ( −28,057-28{,}057 )9,636
Final selectionRemaining a priori exclusions and derived features7,978
Appendix
Table S3: Column count at each stage of construction and exclusion. Counts are columns, including the participant identifier.
GroupEntriesRepresentative fields
Outcome or temporal leakage2,69242018, 42020, 130000–132605, 41272
Diet: single-day 24-hour recall820077, 26068, 105010
Study administration and timing24321711–21771, 20751, 32365–32368
COVID-1920127981–28174, 29156–29161
Employment history8122601, 22602, 22603
Assay and genotyping quality control7222000–22030, 32050–32065, 3061
Appendix
Table S4: A priori feature exclusions by group. Counts are distinct excluded entries, deduplicated across pipeline stages, and are a lower bound on the number of columns removed.
OriginColumnsParent fields
UK Biobank field, one column each6,8543,726
Multiple-choice one-hot indicators1,119112
Derived for this work4—
Total7,9773,838
Appendix
Table S5: Composition of the 7,977-column candidate feature set that remained after the a priori exclusions and entered feature selection.
LibraryParameterValueSource
CatBoostEvaluation metricPRAUCset
 Task typeCPUset
 Threadsall availableset
 Verbositysilent, no on-disk logsset
 Random seedfixedset
 Loss functionLoglosssupplied
Appendix
Table S6: Training configuration of the gradient-boosting models. Set marks a value we passed to the constructor; supplied marks one the library chose, and those follow from the versions listed above. Random seeds were held fixed across runs. MVS, minimal variance sampling.
More in this topic

Clinical Risk Prediction

Clinical risk prediction estimates patients’ future health outcomes from medical records, physiological signals, imaging, and other health data.

See all 217 papers

Explore similar work

CardsList
May 4, 2026

Gradient Boosted Risk Scores

Risk scores are an interpretable and actionable class of machine learning models with applications in medicine, insurance, and risk management. Unlike most computational methods, risk scores are designed to be computed by a human by attributing points to a data sample based on a limited set of criteria. The most common approaches for generating risk scores use linear regressions to estimate the effect of selected variables. We propose a simple and effective approach towards building compact and predictive risk scores. We provide an algorithm based on gradient boosting that is capable of modeling nonlinear effects, along with a C++ implementation with Python and R bindings. Through extensive empirical evaluation on twelve tabular datasets spanning regression, classification, and time-to-event tasks, we show that our method achieves competitive predictive performance while producing substantially more compact scores than regression-based alternatives, with 60% fewer rules for classification tasks and 16% fewer rules for time-to-event tasks on average, compared to AutoScore.

Costa Georgantas, Jonas Richiardi

†Department of Radiology, Lausanne University Hospital and University of Lausanne, Switzerland

Interpretable MLClinical PredictionTabular Prediction
AskPDF
Sep 16, 2026

Machine-Learning Assessment of the Predictive Value of Inflammatory Biomarkers for Cognitive Impairment in an Older Hispanic Adult Cohort

Small clinical tabular datasets require interpretable machine learning because deep learning is often impractical and ensemble models can be difficult to inspect. A key pitfall is that statistical significance does not necessarily imply predictive utility. Using data from the Panama Aging Research Initiative--Health Disparities (PARI-HD) cohort (n=165), we implemented a leakage-safe threshold-likelihood Bernoulli/Categorical Naive Bayes (BNB/CNB) classifier. Within every training fold, each continuous predictor was reduced to a supervised chi-square-derived state, while income entered the model through a categorical likelihood. All data-dependent steps were performed within repeated stratified 10-fold cross-validation with 30 repeats. The demographic baseline achieved a ROC-AUC of 0.630 +/- 0.017. I-309 (CCL1) was the dominant incremental feature, increasing AUC by 0.110, with paired DeLong tests yielding p<0.05 in 100% of repeats. In the pre-specified primary analysis, I-309 produced a fixed-partition DeLong p=0.0018, with robustness assessed across 200 random partitions, where the median p-value was 0.0011. Within the exploratory family of 18 candidate markers, I-309 achieved a Benjamini-Hochberg-adjusted q=0.032 on the frozen partition and satisfied q<0.05 in 85% of random partitions, whereas no other marker demonstrated reliable incremental predictive value. Because the fitted model is an inspectable table of thresholds and class-conditional probabilities, these results identify I-309/CCL1 as an interpretable candidate feature for tabular prediction of cognitive impairment, pending external validation.

Antony Garcia, Gabrielle Britton, Alcibiades Villarreal +3

Worcester Polytechnic Institute, Worcester, MA, USA · Facultad de Ingenier´ıa El´ectrica, Universidad Tecnol´ogica de Panam´a, Panam´a, Panam´a · Centro de Vacunaci´on e Investigaci´on (CEVAXIN), Panam´a, Panam´a +3

Clinical Outcome Prediction
AskPDF
Feb 18, 2026

Construction of a classification model for dementia among Brazilian adults aged 50 and over

To build a dementia classification model for middle-aged and elderly Brazilians, implemented in Python, combining variable selection and multivariable analysis, using low-cost variables with modification potential. Observational study with a predictive modeling approach using a cross-sectional design, aimed at estimating the chances of developing dementia, using data from the Brazilian Longitudinal Study of Aging (ELSI-Brazil), involving 9,412 participants. Dementia was determined based on neuropsychological assessment and informant-based cognitive function. Analyses were performed using Random Forest (RF) and multivariable logistic regression to estimate the risk of dementia in the middle-aged and elderly populations of Brazil. The prevalence of dementia was 9.6%. The highest odds of dementia were observed in illiterate individuals (Odds Ratio (OR) = 7.42), individuals aged 90 years or older (OR = 11.00), low weight (OR = 2.11), low handgrip strength (OR = 2.50), self-reported black skin color (OR = 1.47), physical inactivity (OR = 1.61), self-reported hearing loss (OR = 1.65), and presence of depressive symptoms (OR = 1.72). Higher education (OR=0.44), greater life satisfaction (OR=0.72), and being employed (OR=0.78) were protective factors. The RF model outperformed logistic regression, achieving an area under the ROC curve of 0.776, with a sensitivity of 0.708, a specificity of 0.702, an F1-score of 0.311, a G-means of 0.705, and an accuracy of 0.703. Conclusion: The findings reinforce the multidimensional nature of dementia and the importance of accessible factors for identifying vulnerable individuals. Strengthening public policies focused on promoting brain health can contribute significantly to the efficient allocation of resources in primary care and dementia prevention in Brazil

F. S. Menezes, M. C. F. G. Barretto, E. Q. C. Garcia +2

Instituto de Medicina Integral Professor Fernando Figueira (IMIP) · Future Tech · Universidade Federal Rural de Pernambuco (UFRPE)

Clinical Risk Prediction
AskPDF