Natural Language Clinical Pathways for Automated Coding of Topography and Morphology in Cancer Registries: Leveraging Healthcare Dataflows through the LN-PDTA Algorithm.

Abstract

Background: cancer registries (CRs) are essential tools for oncological surveillance. The accurate coding of topography and morphology (through ICD-O-3 coding), traditionally performed manually, is complex and time-consuming. Artificial Intelligence (AI) offers new opportunities to automate this process, overcoming the limitations of existing algorithms, which often focus only on topography.

Objectives: to develop an AI-based algorithm capable of automatically assigning the combined topography-morphology (topo-morpho) code from a synthetic clinical pathway expressed in natural language (LN-PDTA).

Design: retrospective observational study based on integrated registry and healthcare administrative data. Deterministic record linkage was performed among the CR, administrative databases, and pathology reports (AP), considering clinical events within ±180 days from incidence date. Clinical information (diagnoses, pharmacological therapies, surgical procedures, causes of death, pathology codes) was transformed into chronological clinical tokens concatenated into a single string. The target variable was the combined topo-morpho code assigned by registry coders. An LSTM neural network (embedding=64, hidden=128) was trained to learn token sequences.

Setting and participants: incident cancer cases recorded by the Cancer Registry of the Agency for Health Protection of the Metropolitan Area of Milan in 2017-2018; multiple, benign, and uncertain tumors were excluded.  Main outcome measures: accuracy in the prediction of topography, morphology, and combined topo-morpho. Precision, recall, and F1 score at different confidence thresholds. Secondary analysis for high-incidence cancer sites and identification of the most predictive tokens and information sources.

Results: the dataset included 34,168 cases, split 80:20 into training and test sets. On the test set, the model achieved 89% accuracy for topography prediction, 59% for morphology, and 56% for the combined topo-morpho classification. Performances were better for highly frequent sites (breast 73%; colorectal 61%). For lung and prostate cancers, accuracy for topography reached 94% and 98%, respectively. The most predictive tokens and information sources were identified: pathology reports, mortality data, and surgical procedures for topography; pathology reports, hospital discharge diagnoses, and mortality for morphology.

Conclusions: the LN-PDTA-based neural network approach shows promising results for the most frequent topographies and morphologies, thus enabling automatic coding of a fair number of cases, reducing manual coding time and supporting more efficient cancer registry operations.