Dementia Cohort Identification and SVM Classifier Pipeline for article ‘Machine learning-based prediction of future dementia using routine clinical MRI brain scans and healthcare data’
Parminder Reel 2025-10-29
This is a comprehensive README for the Dementia Cohort Identification and SVM classifier pipeline used in the article:
Machine learning-based prediction of future dementia using routine clinical MRI brain scans and healthcare data
Authors: Parminder Singh Reel, Salim Al-Wasity, Craig Edwards, Smarti Reel, Esma Mansouri-Benssassi, Szabolcs Suveges, Muthu Rama Krishnan Mookiah, Susan Krueger, Emanuele Trucco, Emily Jefferson, Alexander Doney and J. Douglas Steele.
The Dementia Cohort Identification and SVM classifier pipeline was developed under the PICTURES Exemplar-2 Project.
For any queries regarding the code contact the corresponding author at p.s.reel@dundee.ac.uk
This repository documents the methodology used in the study, encompassing both clinical cohort identification and SVM based machine learning based pipeline. The data cohort identification details how dementia and control cohorts are derived from linked healthcare records. The SVM pipeline provides a reproducible framework for model training and evaluation using T1 weighted brain MRI data, demonstrated with the publicly available IXI dataset. These components are presented independently, allowing the ML pipeline to be executed without access to restricted clinical data.
The dementia and control cohorts are identifies using clinical data. This part includes R scripts that perform this step. It uses a previously validated multi-source electronic medical record (EMR) phenotyping algorithm (Doney et al., Diabetes Care, 2019), which integrates hospital admissions (ICD coding), prescribing data (e.g. dementia-specific medications), and mortality records within longitudinal NHS data.
Dementia Cases were selected from GoDARTS and SHARE studies using the following four linked datasets:
- Dispensed prescribing: This dataset includes the detailed longitudinal prescription records of all medications that are dispensed from the pharmacies to the individuals. The drugs specific to dementia i.e. Chapter 4.11 of the British National Formulary (BNF) were identified.
- Death: The Scottish General Register Office (GRO) records information of death certification. The ICD-9 and ICD-10 codes related to dementia were identified in these certifications.
- General hospitalisation admissions (SMR01): The general hospital inpatient and day case discharge information include the individual episode level data. The ICD-9 and ICD-10 codes related to dementia were identified in these records.
- Psychiatric hospital discharge (SMR04): The psychiatric hospital inpatient and day case discharge information include the individual episode level data. The ICD-9 and ICD-10 related to dementia were identified from these records.
For each patient with dementia, the incident date of dementia was evaluated based on the earliest date of occurrence in any of these datasets. As EMR records are independently coded episodes over time, patients could receive codes from different dementia subtypes due to the intrinsic inaccuracy in the clinical coding. A probabilistic model was used to assign a dementia subtype to such patients.
Non-Dementia/Control patients were selected from GoDARTS and SHARE studies and have:
- No dementia specific drugs (Chapter 4.11 of the BNF) in the dispensed prescribing records,and
- No ICD-9 and ICD-10 codes related to dementia in death, general hospital admissions, and psychiatric hospital discharge datasets.
- R (>= 4.0)
- Packages:
dplyrstringrDBIodbc
The pipeline identifies patients with and without dementia using clinical records. The workflow is split into two main parts:
- Dementia cohort construction
- Control cohort construction
Purpose
-
Identify dementia patients using ICD codes from:
- SMR01, hospital admissions
- SMR04, psychiatric records
- GRO, death records
- Prescribing data
-
Apply a probabilistic model to assign dementia subtype:
- Alzheimer’s disease
- Vascular dementia
- Other dementia
- Unspecified dementia
Outputs
All_Dementia_patients.csv
Purpose
- Remove all dementia patients
- Identify individuals with no dementia records
- Link to imaging metadata
- Extract follow up information from clinical datasets
- Compute latest date of follow up
Outputs
All_Non_Dementia_patients.csv
The pipeline uses:
- Clinical databases
- SMR01, hospital admissions
- SMR04, psychiatric admissions
- GRO, death records
- Prescribing data
Note, current scripts connect to an SQL server. In other environments, these should be adapted to CSV inputs.
This SVM Pipeline uses SPM and MATLAB. For demonstration purpose, it uses publicly available T1 Brain scans from IXI database for the entire workflow and describes configuration, outputs, and how to run the pipeline.
- MATLAB R2021b or later
- SPM12 (Statistical Parametric Mapping toolbox)
- Parallel Computing Toolbox (for distributed jobs)
- IXI Dataset (T1-weighted MRI scans)
Below is a detailed, practical breakdown of each pipeline stage. Each step includes the purpose, rationale, and commands to run the step.
Purpose & Rationale
The data related to the classification results presented in this article can only be
accessed within the HIC TRE subject to ethical and governance approvals.
Therefore, publicly accessible T1 Brain Scans from IXI data is used to
demonstrate the functional pipeline. Also, create a single master table
that links IXI metadata to MRI file paths and produces the binary
classification labels using patient age. This is critical because
downstream cross-validation and stratification depend on accurate
mapping between subjects and images.
Download T1 Brain Scans from (https://brain-development.org/ixi-dataset/)
Command
srun -n1 -N1 /usr/local/MATLAB/R2021b/bin/matlab -nosplash -r "Rseed= 1; setup_IXI_data"Purpose & Rationale
Implement nested cross-validation to provide an unbiased estimate of
generalization and to separate model selection (inner loop) from
performance estimation (outer loop).
Command
srun -n1 -N1 /usr/local/MATLAB/R2021b/bin/matlab -nosplash -r "Rseed= 1; Step1"Purpose & Rationale
Produce job driver tables that enumerate all independent combinations of
outer/inner folds, p-thresholds, and hyperparameters. These tables
enable parallel dispatch to cluster nodes or local workers and act as
reproducible experiment manifests.
Command
srun -n1 -N1 /usr/local/MATLAB/R2021b/bin/matlab -nosplash -r "Rseed= 1; Step2_alt"Purpose & Rationale
Standardize all images to a common anatomical space (MNI) and segment
tissue classes. SPM’s unified segmentation + normalization or
shoot/DARTEL produce images that are comparable across subjects and
suitable for voxelwise statistics. Perform voxel-wise statistical tests
to detect voxels showing significant group differences. Masks derived
from these p-values reduce dimensionality and help focus the classifier
on biologically plausible regions.
Command
srun -n25 -N25 /usr/local/MATLAB/R2021b/bin/matlab -nosplash -r "Rseed= 1; Step3_alt"Purpose & Rationale
Train SVM classifiers on masked voxel features and evaluate on held-out
inner folds to estimate performance of each hyperparameter combination.
Command
srun -n128 -N64 /usr/local/MATLAB/R2021b/bin/matlab -nosplash -r "Rseed= 1; Step4_alt"
srun -n1 -N1 /usr/local/MATLAB/R2021b/bin/matlab -nosplash -r "Rseed= 1; Step4a"Purpose & Rationale
Aggregate inner-fold results to select the best hyperparameter
configuration for each outer fold without tapping into outer test data.
Command
srun -n5 -N5 /usr/local/MATLAB/R2021b/bin/matlab -nosplash -r "Rseed= 1; Step5"Purpose & Rationale
Use the chosen hyperparameters to train a final model on the full outer
training set and evaluate on the outer test set to estimate
generalisation.
Command
srun -n5 -N5 /usr/local/MATLAB/R2021b/bin/matlab -nosplash -r "Rseed= 1; Step6_alt"Purpose & Rationale
Validate that all jobs completed correctly, aggregate fold-level
results, compute summary statistics, and prepare a report for
publication or further analysis.
Command
srun -n1 -N1 /usr/local/MATLAB/R2021b/bin/matlab -nosplash -r "Rseed= 1; Step7_alt"The pipeline records a comprehensive set of performance metrics for each evaluation. Each metric is saved per job and aggregated across folds.
Per-model / per-fold metrics
- ACC — Accuracy (correct predictions / total observations)
- AUC — Area Under ROC Curve (discrimination performance)
- SENS — Sensitivity / True Positive Rate
- SPEC — Specificity / True Negative Rate
- PPV — Positive Predictive Value
- NPV — Negative Predictive Value
- NUM_OBS — Number of observations in test set
- Confusion matrix entries —
CM11,CM12,CM21,CM22
Below are the most important configuration variables controlling the
experiment. Keep track of them in config_experiment.m.
seed— random seed for reproducibility (e.g.,42)kFolds_outer— number of outer folds (e.g.,5)kFolds_inner— number of inner folds (e.g.,3)p_Thres_range— p-value thresholds for mask creation (e.g.,[0.01 0.001 0.0001])C_value_range— SVM regularization constants (e.g.,[0.01 0.1 1 10])K_value— kernel scale parameter when using Gaussian/RBFspmpath— path to SPM12 installationimagespath— path to preprocessed imagesttestoutpath,driverpath,outertraintestpath,innertraintestpath— output directoriesmin_p,max_p,min_C,max_C— ranges for final summary bookkeeping
A consistent directory & filename convention is used so downstream analysis is straightforward.
Top-level output directories
ttestout/— p-matrix and binary masks ({outer}_{inner}_ttest2_result.mat,{p}_binarymask.mat)innertraintest/— inner fold training/validation CSVs and per-worker logsoutertraintest/— outer fold evaluation CSVs and predictionssplits/— outer/inner split CSVsdriver/— driver CSVs used to distribute workruntime_log/— step-wise runtime logs (Step1_runtime.txt, etc.)
Important file formats
.mat— MATLAB data objects (p-matrix, masked volumes, saved SVM models).csv— human-readable driver & results tables.nii— neuroimaging volumes (preprocessed tissue maps).txt— runtime or completion flags
Example final artifacts
outer_fold_summary.csv— final performance summary across outer foldsInner_loop_validated.txt— indicates all inner jobs completedfinal_masks/— best masks for each outer fold (for interpretation)predictions/— subject-level predicted labels & decision scores
This work was funded by the Medical Research Council (MRC) MICA Programme Grant: InterdisciPlInary Collaboration for efficienT and effective Use of clinical images in big data health care RESearch: PICTURES (MR/S010351/1).
This code is licenced under the MIT License.
