The software pipeline is composed of three main modules: Cohort Planner, Cohort Extractor and Cohort Labeller (as shown in Fig. 1). Integration of these subcomponents enhances the overall functionality and effectiveness of the pipeline, enabling clinical researchers to efficiently extract, label, and assess patient data for research purposes while ensuring data quality and reliability.
3.1 The Cohort Planner Module
This module allows clinical researchers to estimate the number of potential patients available for analysis based on selected inclusion and exclusion criteria. It assists in planning data extraction by providing an estimate of the sample size available for analysis.
CohortPlanner integrates practical statistical measures to enhance cohort utility. These measures include power calculations to determine sample size adequacy for planned cohorts and estimates of deviations from expected norms to assess the completeness and reliability of EMR data.
3.2 The Cohort Extractor Module
The main component is a module that streamlines the process of cohort assembly.
This is achieved by performing the automatic extraction of raw imaging data retrieved from the machine databases, as well as AI-assisted extraction of retinal biomarkers. It performs two main functionalities: uploading a cohort, and downloading it (as shown in Fig. 2).
The process of “uploading” a cohort starts with the identification of patient groups based on specific diagnostic criteria. This can either be done directly via patient identifiers, or indirectly via selection criteria, such as age, gender, and more. Concurrently, patient consent for data usage is verified by querying the General consent database, ensuring compliance with privacy regulations and legal provisions on research involving human beings (HRA). Upon identification, the corresponding raw imaging data is gathered from a centralised image pool. Following this, imaging data is uploaded to an instance of Discovery® software by RetinAI. After having retrieved the resulting analysis, one can perform automated, integrated post-hoc analysis in order to validate or augment Discovery® ‘s data. At present, a machine learning model is used to further label gathered eye fundus images.
The process of “downloading” a cohort involves the user specifying a configuration to define which image types and biomarkers are to be considered. The cohort is then downloaded from Discovery® to the local file system in a selective manner, based on the aforementioned configured study data.
3.3 The Cohort Labeller Module
Designed to facilitate expert labelling of the extracted patient dataset, this software module enables clinical researchers and clinicians to systematically assess each patient's characteristics and label them according to the cohort they belong to or identify relevant biomarkers. These labelled datasets are crucial for training AI algorithms and can be reused for future studies.
Cohort Labeller is an interactive tool that facilitates dynamic data interaction. This platform enables the visualisation of specific scans, segmentation of chosen parameters, and the distribution of crucial fluids (IRF, SRF) via histogram overlays. It has increased data labelling efficiency, allows for the integration of treatment histories within the analysis, and supports the selection of displayed parameters, significantly streamlining the interpretative process for medical professionals. Cohort Labeller increases the usability of longitudinal data for ophthalmological research, accentuating Cohort Builder's utility in fostering data-driven insights in clinical settings.
3.4 The EMR Health Checker Module
This software module, separate from the rest of the pipeline, provides indicators of the reliability of certain fields within the Electronic Medical Record (EMR) database. It offers a description of the data quality, helping to estimate the reliability of the extracted cohorts.
The main indicators are extracted from a registry of medical consultations, diagnoses, treatment plans, and surgical interventions, before being combined and evaluated. These indicators can also be utilised for training assistant clinicians and improving clinical practice.
3.5 Use Cases
The Cohort Builder pipeline has played a crucial role in the formation of patient cohorts with specific baseline characteristics for a range of ophthalmology projects. This summary highlights its application across various studies, outlining the creation of cohorts for patient groups identified by particular ocular conditions under the Swiss Ophthalmic Imaging Network (SPHN) initiative (Lawrence, Selter, and Frey 2020).
Key applications of Cohort Builder included the creation of the following retrospective study cohorts. More information is available on the SOIN website at https://sphn.ch/network/projects/completed-projects_tiles/project-page_soin.
1. Grading of Uveitis Inflammation: A cohort to study inflammation in uveitis patients (3312 eyes; 534 patients; Fluorescein Angiography [FA]) (Amiot et al. 2023).
2. Comorbidity of Glaucoma and Diabetes: A cohort to compare patients with primary open-angle glaucoma (POAG) without other comorbidities, and patients with both POAG and T2D (77 patients; Optical Coherence Tomography [OCT]).
3. Central Serous Chorioretinopathy Recurrence: A cohort to compare recurrent versus non-recurrent central serous chorioretinopathy (CSCR) patients (344 eyes, 255 patients; OCT and Fundus Autofluorescence [FAF]). The dataset supported both statistical and machine learning analyses to identify factors influencing recurrence, as well as deep learning predictions for recurrence.
4. Retinal Layer Thicknesses in Diabetic Patients: A cohort study to measure retinal layer thicknesses across several diabetic patient subgroups was conducted, including various stages of non-proliferative diabetic retinopathy (NPDR) and proliferative diabetic retinopathy (PDR), alongside control groups (127 NPDR type I patients, 49 NPDR type II, 19 NPDR type III, 176 PDR, 75,751 controls; [OCT]).
5. Geographic Atrophy Identification: A cohort capturing geographic atrophy in wet age-related macular degeneration (100 patients; [OCT]), including equal numbers of cases and controls.
6. Photoreceptor Imaging using Adaptive Optics: A cohort of cellular-resolution images of cone photoreceptors captured using Adaptive Optics Scanning Laser Ophthalmoscopy (AOSLO) technology (43 patients; [AOSLO] and [OCT]).
7. OphtalmoLaus cohort: A cohort including diverse imaging modalities (2208 patients; [OCT], [OCT Angiography], [Color Fundus Images (CFI)], and [iris photographs]). The Cohort Builder pipeline facilitated the extraction of 46 independent eye parameters, including layer thicknesses and disease biomarkers.
8. Ocular Genomics: A cohort comparing OCT-derived parameters from the UK biobank (N=88,248; [OCT] and [CFI]) for genomic analysis (Bergmann et al. 2023; Tomasoni et al. 2023).