Tumor data extraction from unstructured medical reports: application of an artificial intelligence-enhanced tool in the Eastern Switzerland Cancer Registry

Abstract

Background

Population-based cancer registries rely heavily on the manual coding of unstructured medical reports, a process that is both time-consuming and resource-intensive. The number of cases and reports is increasing and simultaneously qualified coding personnel are rare, making it challenging to maintain timely and high-quality cancer registration.

Materials and methods

Between March and September 2025, we implemented and evaluated an artificial intelligence (AI)-enhanced tool, a modular Python-based application, for tumor data extraction from 82 384 medical reports received by the Eastern Switzerland Cancer Registry. The tool, based on fine-tuned, pretrained German-language bidirectional encoder representations from transformers models, processed reports from >200 institutions, and provided an output file that can be directly integrated into the Swiss cancer registry software NICERStat-KRG (Cantonal Cancer Registries, Switzerland).

Results

For 10 739 reports (13%), the incidence date, International Classification of Diseases (ICD) 10 code, and ICD-O-3.2 topography, morphology, and behavior showed complete agreement with the corresponding tumor records in our database. Remaining reports were validated and registered by coding staff. Across core cancer registry variables, the tool achieved moderate-to-high agreement with the coding by our staff, with exact match rates of 19%-53% and partial matches up to 82%. It also correctly classified 82% of previously ignored reports as irrelevant. Despite a 27% reduction in coding staff resources, overall productivity increased markedly compared with 2024.

Conclusions

These findings demonstrate that a locally deployed, privacy-preserving AI-enhanced tool can substantially improve productivity and data completeness in population-based cancer registration. The results support the potential of AI-assisted workflows to facilitate routine registry operations under increasing workload demands.

Highlights

• •Local, privacy-preserving AI significantly improves cancer registry productivity.
• •AI increases data completeness in population-based cancer registration.
• •Productivity improved despite reduced staffing and rising workload.
• •Tool supports sustainable cancer registration under real-world constraints.
• •Findings are relevant to cancer registration, epidemiology, and health informatics.

Introduction

Population-based cancer registration provides essential data for cancer control but manual processing of unstructured medical reports remains a key challenge.1, 2, 3 In Switzerland, the implementation of the National Cancer Registration Act in 2020 led to obligatory reporting of all tumors [International Classification of Diseases (ICD)-10: C00-C43, C45-C97, D00-D03, D05-D07, D09, D32, D33, D35.2, and D37-D48] and a steady annual increase in the number of reports per tumor case. Given the scare availability of qualified staff, innovative approaches are required to ensure timely and complete registration.⁴

Artificial intelligence (AI) and natural-language processing (NLP) have shown potential to extract structured clinical information from free-text pathology or radiology reports.5, 6, 7 However, real-world implementation in national or regional cancer registries remains limited. To the best of our knowledge, this is the first study reporting on how a real-world implementation of an AI-enhanced tool improved productivity in a population-based cancer registry. We report the design and the method of an AI-enhanced, Python-based data-extraction framework that was implemented into the workflow of the Eastern Switzerland Cancer Registry and how it affected productivity in registering and editing tumor cases given the constraints of scarce staffing.

Materials and methods

Data sources and study period

Between 24 March and 1 September 2025, an AI-enhanced tool was deployed in routine operations at the Eastern Switzerland Cancer Registry. During this period, the system processed 82 384 unstructured medical reports that had been issued and/or delivered to the registry between January and November 2024.

The reports originated from >200 institutions, including acute care hospitals, pathology and cytology laboratories, radiology institutes, and office-based general practitioners and specialists. Document types included pathology, cytology, radiology, tumor board, and hospital discharge reports. All documents were received in a portable document format (PDF) and represented the standard input material for cancer case ascertainment and coding at the registry.

System architecture and AI models

The system was implemented as a modular Python application and integrated into the existing IT infrastructure of the cancer registry (Figure 1). The workflow consisted of the following components:

• •Document ingestion and preprocessing: Incoming PDF documents were automatically ingested and converted to machine-readable text using optical character recognition where required. The resulting text was serialized into structured JavaScript Object Notation (JSON) files, preserving document identifiers, page structure, and metadata. Rule-based preprocessing and normalization were subsequently applied, including removal of headers, footers, duplicated content, and other noninformative text elements, before model inference and information extraction.
• •Document classification and model inference: Fine-tuned German Medical bidirectional encoder representations from transformers models were applied to classify document types and identify tumor-related content. The models were trained and fine-tuned using a manually annotated corpus of 178 667 clinical reports from the Eastern Switzerland Cancer Registry, representing a mix of pathology, cytology, radiology, tumor board, and discharge reports from 2020 to 2023, after the Cancer Registry Law came into force. The AI-enhanced tool used a master Excel file containing all relevant tumor variables as input, enabling structured extraction of both clinical and administrative data directly from the JSON files. Annotation was carried out by experienced cancer registry staff according to internal, National Association of Cancer Registries and international standards of the European Network of Cancer Registries, the International Agency for Research on Cancer, and the Union for International Cancer Control, ensuring high-quality ground truth for model training.

Figure 1.

Flowchart diagram of the AI-enhanced tool pipeline.

Model inference focused on extracting key clinical entities and attributes relevant for cancer registration, including ICD-10 codes, ICD-O-3.2 topography and morphology, behavior, histological grade, laterality, and incidence date.⁸

• •
  Information extraction: The system extracted predefined variables relevant for cancer registration, including

  • °
    Core registry variables: incidence date, ICD-10 diagnosis, ICD-O-3.2 topography and morphology, laterality, behavior, and histological grade [including grading systems such as Greasmann, Elston–Ellis, and World Health Organization Classification of Tumors of the Central Nervous System (WHO CNS)].
  • °
    Administrative variables: source and type of medical report, report date, referring physicians, institutions and hospitals, and patient notification date.
• •Output generation and quality control: All extracted information was written to structured JSON files and summarized in a comma-separated values (CSV) file. Visualization and logging components were implemented to flag missing, conflicting, or low-confidence outputs for quality control and downstream review.
• •
  Coding agreement was defined as follows:

  • •
    Document level

    • °
      Document-level accuracy was defined as the proportion of reports in which all predefined core registry variables (incidence date, ICD-10 diagnosis, ICD-O-3.2 topography, morphology, behavior, laterality, and histological grade) were correctly extracted without requiring manual correction.
  • •
    Variable level

    • °
      Exact match: The AI-extracted variable exactly corresponds to the registry staff-validated registry entry (e.g. identical ICD-10 code, ICD-O-3.2 morphology, or incidence date).
    • °
      Partial match: The AI-extracted variable is partially correct, for example, matching topography at the first three-digit level and morphology at the first four-digit level or an incidence date within the same month and year as the registry staff-validated date. Such partial matches are generally sufficient for internal cancer registry workflows, including case identification, workload prioritization, and the preparation of annual registry statistics and reports. However, they do not meet the requirements for final registry coding and were therefore flagged for registry staff review and completion.

Integration into registry workflow and registry staff validation

The AI-enhanced tool was integrated as a decision-support system rather than as a fully autonomous coding solution. The extracted data were imported in a CSV format into the national cancer registry software NICERStat-KRG (National Institute for Cancer Epidemiology and Registration, MySQL-based), where they were made available to trained cancer registry staff within the routine case processing interface.

Registry staffs reviewed, validated, corrected where necessary, and finalized all extracted variables in accordance with established internal, national, and international coding rules. No case was registered without registry staff verification. The AI output therefore served to prepopulate registry fields and support case ascertainment, while full responsibility for final coding remained with registry staff.

Unit of analysis. The primary unit of analysis was the individual medical report (document). During routine cancer registration, multiple reports (e.g. pathology, radiology, tumor board, and discharge reports) may be received for the same patient, each containing different subsets of clinical information. The AI-enhanced tool processed each report independently and extracted all predefined cancer registry variables available within that document. Consequently, performance was evaluated at two complementary levels: (i) document level, assessing whether all relevant variables contained in an individual report were correctly extracted, and (ii) variable level, assessing agreement for each individual registry variable across all processed reports.

Evaluation of productivity and staffing impact

Productivity was assessed by comparing operational indicators during the AI implementation period (24 March to 1 September 2025) with the same calendar period in 2024, before the introduction of the without AI assistance.

The following indicators were evaluated:

• •Number of processed medical reports.
• •Number of newly registered tumor cases.
• •Number of automatically generated questionnaires for missing patient notifications.
• •Staff availability, measured in full-time equivalents (FTEs).

Baseline conditions in 2024 represented fully manual processing, in which registry staffs reviewed and coded all incoming reports without AI-based pre-extraction. Changes in productivity were interpreted in relation to staff availability and case volume, reflecting real-world registry operations rather than controlled experimental conditions.

Results

Of 82 384 processed reports, 2403 (2.9%) were unreadable PDFs and some 100 were in non-German languages. The AI-enhanced tool was thus not able to process these files. For 65 572 (80%) reports, relevant tumor information was extracted and were classified as relevant.

All reports, whether or not AI output was available, were imported into NICERStat-KRG. For 10 739 reports (16% of processed reports or 13% of total reports), the incidence date, ICD-10 code, and ICD-O-3.2 topography, morphology, and behavior completely matched the corresponding tumor records already in our database. For these reports, no manual reabstraction or recoding of core variables was required. This subset of reports corresponds to an estimated annual reduction of 70% of a FTE in manual coding workload, reflecting time saved on data entry and coding rather than elimination of registry staff review.

For the remaining 71 645 reports, extracted tumor variables (incidence date, ICD-10, ICD-O-3.2 topography and morphology, laterality, behavior, and histological grade, including Greasmann, Elston–Ellis, and WHO CNS systems), as well as administrative data (patient notification date and details on the source, type, and date of reports) were fully reviewed, validated, and processed by registry staff.

The performance of the AI-enhanced tool for extracting core cancer registry variables from unstructured medical reports was evaluated across several core cancer registry variables (Table 1). For incidence dates, exact matches were achieved in 19% of cases, increasing to 48% when considering month/year-level matches. For ICD-O-3.2 topography, exact matches occurred in 26% of cases, with 47% matching at the first three digits. ICD-O-3.2 morphology matched exactly in 33% of cases and at the first four digits in 35%. ICD-O-3.2 grading exact matches were observed in 26% of records, while ICD-10 diagnoses showed 15% exact matches and 20% matches using only the first three digits of the ICD-10 code. Laterality had the highest exact match rate at 53% and specialized grading systems (Greasmann, Elston–Ellis, and WHO CNS) matched exactly in 48% of cases. Among medical reports that had been ignored by registry staffs (n = 18 235), the AI-enhanced tool correctly classified 82% as not relevant, while incorrectly classifying 22% as relevant.

Table 1.

Accuracy of AI-enhanced tool data extraction for core cancer registry variables

Core variables	No.	Accuracy (%)
1. Incidence date
Exact match	11 988	19
Match by month/year level	30 584	48
2. ICD-10 diagnosis
Exact match	9375	15
Match of the first three digits	12 673	20
3. ICD-O-3.2 topography
Exact match	16 682	26
Match of the first three digits	29 881	47
4. ICD-O-3.2 morphology
Exact match	21 075	33
Match of the first four digits	22 181	35
5. Histological grade
Exact match	16 291	26
6. Grading systems (Greasmann, Elston–Ellis, and WHO CNS)
Exact match	7275	48
7. Laterality
Exact match	33 422	53
8. Ignored medical reports by coding staff
All ignored reports	18 235
Agreement with the AI-enhanced tool	14 971	82
Disagreement with the AI-enhanced tool	3264	22

WHO CNS, World Health Organization central nervous system.

Despite a 27% reduction in available staffs measuring full-time equivalent compared with the period without AI assistance (410% versus 560%), the registry achieved substantial productivity gains: newly registered tumors increased by 69%, the number of processed documents rose by 92%, and automatically generated questionnaires for missing patient notifications increased by 33% (Table 2).

Table 2.

Productivity and case registration before and after implementation of the AI-enhanced tool

Indicators	24 March to 1 August 2024	24 March to 1 August 2025	Difference (%)
Number of new registered tumors	4351	7371	69
Number of medical reports	31 113	59 592	92
Number of generated questionnaire	3384	4508	33
Total registry staff availability (FTE) (%)	560	410	−27

FTE, full-time equivalent.

Discussion

This study demonstrates the feasibility and real-world impact of integrating an AI-enhanced information extraction system into the routine workflow of a population-based cancer registry. Consistent with findings from previous NLP-based cancer registry and clinical abstraction studies, the system substantially increased processing efficiency and case throughput while maintaining mandatory human validation of all registry data.6, 7, 8 Similar productivity gains have been reported in registry-adjacent settings using rule-based or transformer-based NLP approaches; however, many of these studies were conducted in experimental environments or focused on single data sources such as pathology reports. In contrast, the present implementation operated under routine registry conditions and processed heterogeneous clinical reports from multiple institutions, reflecting the operational complexity of population-based cancer registration.

A key strength of the approach lies in its hybrid model of AI-assisted workflow and human oversight. Rather than replacing human coders, the AI tool functioned as a decision-support system that prepopulated registry variables and supported case ascertainment, while final responsibility for coding and validation remained with trained registry staff. This balance is essential in cancer registration, where clinical nuance, longitudinal case interpretation, and adherence to national and international coding standards cannot be fully automated. The findings underscore that efficiency gains in this context are driven by reductions in manual data entry and repetitive coding tasks, not by the elimination of expert review.

From a data governance perspective, the system was deployed entirely on local infrastructure, avoiding cloud-based processing and ensuring compliance with Swiss data-protection and cancer registry legislation. This architectural choice is particularly relevant for population-based registries operating under strict legal mandates and public trust obligations. At the same time, sustainable use of AI in registry settings requires ongoing model maintenance, monitoring for performance drift, and periodic retraining to reflect evolving clinical terminology, reporting practices, and coding standards. Simultaneously, the implementation of AI may also introduce or amplify human biases in coding and validating. Several limitations should be acknowledged. The models were trained and evaluated using data from a single regional cancer registry and primarily German-language reports, which may limit generalizability to other linguistic, institutional, or health care contexts. Performance may vary in settings with different document structures, reporting conventions, or levels of data completeness. In addition, the current tool has limitations in processing image-based PDF documents, non-German reports, and linking multiple reports to a single tumor episode. These factors may introduce institution- or language-specific biases that warrant careful monitoring.

Future development will focus on multilingual model expansion, improved optical character recognition for scanned documents, and enhanced longitudinal case aggregation across reporting sources. Scaling AI-enhanced cancer registration will also require targeted training for registry personnel, clear governance frameworks for algorithm updates, and continued alignment with evolving regulatory and ethical standards. When implemented transparently and with sustained human oversight, AI-supported extraction systems have the potential to strengthen population-based cancer surveillance, improve data timeliness, and support evidence-based cancer control at the national and international level.