Abstract 4087: Developing a standardized framework for curating oncology datasets generated by manual abstraction and artificial intelligence
Bibliographic record
Abstract
Abstract Background: The widespread uptake of electronic health records (EHRs) has made the creation of custom, real-world datasets for research more feasible. As a result, multiple research datasets with overlapping populations are often generated, using different methodologies, and frequently siloed within and between research groups, limiting the scope of the data’s use. Currently, there is no standard for collating and evaluating such data. Using existing lung oncology datasets, we developed an approach to determine optimal methods of combining and curating clinical data from different sources. Methods: Two separate study datasets containing data for lung cancer patients diagnosed and/or treated within Princess Margaret Cancer Centre (PM, Toronto) were investigated. Study 1 manually abstracted clinical data for 1,990 patients, first seen at PM between 2014-2016; Study 2 leveraged the artificial intelligence engine, DARWEN™, to extract clinical data directly from EHRs for 4,466 patients, diagnosed between 2014-2018. Each dataset was individually assessed for internal consistency before comparing the overlapping population (Test Group, n=1892) to identify, investigate, and resolve differences. Patterns of data extraction performance were evaluated to define optimal methods for combining datasets and informing future data collection. Herein, epidermal growth factor receptor (EGFR) mutation status is used as an illustrative example. Results: Study 1 and 2 had similar distributions of clinicodemographic data and frequency of EGFR mutations. The Test Group had 100% agreement for date of birth, and >99% agreement for sex, with all discrepancies resulting from human error in Study 1. The Test Group had a 98% agreement for EGFR positivity and 98-99% agreement for specific exon mutations. Of the 106 disagreements for specific mutations, 50% (n=53) were due to Study 1 human error. Study 2 prioritized specificity over sensitivity for biomarker extraction, resulting in more false negatives (25% of errors, n=26). As DARWEN™ only extracted EGFR data from pathology reports, 18% (n=19) of discrepancies were due to lack of access to relevant information captured elsewhere in patients’ EHRs. Adjudicators could not resolve the remaining 7% of disagreements (n=8). Conclusions: By comparing overlapping datasets, the strengths and weaknesses of each study design and extraction methodology were identified. This process demonstrated the effectiveness of artificial intelligence for extracting accurate patient-level clinicodemographic and mutation status data from EHRs, and the value of targeted manual chart review. Our approach provides a roadmap for leveraging existing clinical datasets to their fullest potential, which is relevant across diverse data extraction methods and study designs. Citation Format: Benjamin M. Grant, Aein Zarrin, Luna Zhan, Rami Ajaj, Lina Darwish, Khaleeq Khan, Devalben Patel, Kaitlyn Chiasson, Karmugi Balaratnam, Maisha T. Chowdhury, Amir-Arsalan Sabouhanian, Joshua Herman, Preet Walia, Evan Strom, Catherine Brown, Miguel Garcia-Pardo, Sabine Schmid, Christopher Pettengell, Erin L. Stewart, Geoffrey Liu. Developing a standardized framework for curating oncology datasets generated by manual abstraction and artificial intelligence [abstract]. In: Proceedings of the American Association for Cancer Research Annual Meeting 2022; 2022 Apr 8-13. Philadelphia (PA): AACR; Cancer Res 2022;82(12_Suppl):Abstract nr 4087.
Fetched live from OpenAlex and de-inverted. Abstracts are not stored in this database: the inverted indexes are 8.6 GB of the frame’s 9.3 GB of text, and the host has 13 GB free.
How this classification was reachedexpand
Full frame machine prediction
Teacher imitationNot calibrated prevalence, not ground truth. Human validation pending. The Gemma side is a direct model label for every work in the frame, read from the title-only record. The Codex side is a classifier learned from the 10,348 direct Codex labels and calibrated to design-weighted sample rates; fields without enough sample support carry no Codex call. Candidate is the union of the two sides; consensus is their intersection. These outputs are machine_predicted_unvalidated and are not human labels.
Distilled classifier scores by category (both heads)
| Category | Codex | Gemma |
|---|---|---|
| Metaresearch | 0.050 | 0.076 |
| Meta-epidemiology (narrow) | 0.003 | 0.002 |
| Meta-epidemiology (broad) | 0.002 | 0.006 |
| Bibliometrics | 0.013 | 0.008 |
| Science and technology studies | 0.002 | 0.002 |
| Scholarly communication | 0.010 | 0.008 |
| Open science | 0.006 | 0.014 |
| Research integrity | 0.002 | 0.004 |
| Insufficient payload (model declined to judge) | 0.004 | 0.005 |
Machine scores (provisional)
The two teacher heads of the student model, read on this work. A score orders the frame for review; it never asserts a category, and the validation status ships verbatim with every row.
Baseline scores from an immature model (maturity gate not passed, 7 training rounds). Scores rank; they never assert a category.
score_only:v0-immature-baseline · verbatim from the scoring run: score_only means the number may rank works, and no category label ships from itClassification
machine, unvalidatedMachine predicted; a candidate call from one source (direct Gemma or distilled Codex), not a consensus.
How this classification was reached, model by model and score by score, is at the end of the page under "How this classification was reached".