Let AI handle the data work. Let researchers focus on research.

How AI-assisted cohort identification and clinical data extraction reduced months of manual work to a couple of weeks in a retrospective NSCLC study.

The challenge: turning clinical records into research data

Retrospective research often starts with a time-consuming task: finding the right patients and collecting the data needed for analysis. Researchers may need to review hundreds or thousands of clinical records. They search through physician notes, radiology and pathology reports, treatment records, and other documentation. They then identify eligible patients and manually extract the required clinical variables. This can take months.

A retrospective study conducted by VITO, UAntwerpen, and ZAS (Ziekenhuis aan de Stroom) evaluated an AI-assisted workflow, developed by Monsana, for cohort identification and clinical data extraction in non-small cell lung cancer (NSCLC).

The result: a process that previously took several months was completed in a couple of weeks.

"AI-based screening and extraction clearly delivers time savings compared to a fully manual approach."

— Renée Spoelman, PhD student, UAntwerpen

The study

The study included therapy-naïve patients with Stage IV NSCLC who received first-line immunotherapy and had a documented treatment response.

The objective was twofold: (1) to identify the eligible patient cohort from a larger screened population, and (2) to extract the clinical variables required for subsequent analysis (spanning patient and diagnosis characteristics, disease characteristics, and treatment and response over time) from free-text clinical notes and reports.

An AI-assisted workflow for research

The workflow consisted of two steps. First, AI-assisted screening was used to identify the patients who met the predefined eligibility criteria. From the 1,154 patients initially screened, 180 patients were identified as eligible for the final study cohort. Second, the relevant clinical data were extracted from the records of these 180 patients, providing the dataset required for the subsequent analysis.

Instead of manually reviewing all >1,000 patient records from start to finish, the AI-assisted workflow screened patient records against the predefined eligibility criteria, identified patients who met these criteria, and extracted relevant clinical variables from their unstructured documentation. Researchers could then focus their attention where it was most needed: reviewing and validating the results rather than collecting every data point manually.

Figure 1.  Patient flow diagram for retrospective cohort identification in the VITO/ZAS NSCLC study. Percentages in the main flow are relative to the total screened cohort; percentages in exclusion boxes are relative to the preceding stage. The first 250 records were processed as a pilot batch to validate the screening and classification approach; the remaining 904 records were processed using the validated criteria.

Human oversight and privacy in AI-assisted research

Each evaluation made by the algorithm is linked to the corresponding sentence in the source clinical record, allowing reviewers to verify individual values against source documentation without re-reviewing the complete chart. This traceability supports a targeted quality-control approach, in which review can focus on specific inclusion criteria, exclusion decisions, or extracted variables rather than requiring exhaustive review of every record.

Patient privacy is protected throughout the entire process. Clinical data are handled in accordance with applicable GDPR and data-protection requirements, with appropriate safeguards in place to protect the confidentiality and security of patient information.

Comparative efficiency: months of manual work to only a couple of weeks

A direct comparison is possible because the same researcher conducted the exact same data collection at another hospital that did not use the AI-assisted Monsana workflow. At that site, the same PhD researcher manually identified the NSCLC cohort and extracted the required data by reviewing patient records chart by chart, a process that took several months to complete. In contrast, at ZAS, the AI-assisted workflow automatically excluded 84% of patients, reducing the number of records requiring manual review from 1,154 to only 180. Consequently, data collection at ZAS was completed within just a few weeks, demonstrating a substantial reduction in the manual effort and time required for cohort identification and data extraction.

“Without the help of Monsana’s AI-assisted patient selection tool, it would not have been possible to include enough patient data in our research project. This tool helps us to focus on the actual research part of the project!”

— Inge Mertens, R&D, VITO

The key benefit is not simply faster data collection. It is giving researchers their time back.

Instead of spending months searching through records and copying information into datasets, researchers can spend more time on study design, analysis, interpretation, and generating new insights.

Conclusion

Retrospective research still needs clinical expertise and validation. AI does not replace that. What it can replace is much of the repetitive manual work.

In this study, a process that previously took months was completed in a couple of weeks. AI-assisted screening and data extraction helped turn clinical records into a research-ready dataset faster, while clinicians remained in control of eligibility and classification decisions.

The goal is simple: less time collecting data, more time doing research.

Institutions conducting similar retrospective studies are welcome to contact Monsana to discuss their research needs.