Handling uncertainty using features from pathology: opportunities in primary care data for developing high risk cancer survival methods

Goce Ristanoski, Jon Emery, Javiera Martinez-Gutierrez, Damien McCarthy, Uwe Aickelin

TL;DR

Researchers have developed a method to predict 2-year survival for lung cancer patients using only common primary care blood tests, even when the data is irregular or incomplete. By engineering features that capture "out-of-range" results and borderline "soft range" values, the study demonstrates that routine pathology history contains significant prognostic signals that could help GPs prioritize high-risk patients.

What Problem Does This Paper Address?

General Practitioners (GPs) are the front line of cancer detection, yet they face a massive "signal-to-noise" problem. Cancer symptoms are often non-specific, and the disease has a low prevalence in primary care settings.

Furthermore, primary care data is notoriously "messy." Unlike controlled clinical trials, real-world medical records are plagued by uncertainty: patients miss appointments, tests are taken at irregular intervals, and data is often missing. This paper tackles the challenge of using this incomplete, "uncertain" pathology history to identify high-risk lung cancer patients—specifically those unlikely to survive two years post-diagnosis—to provide better decision support for immediate clinical action.

Methods

The authors focused on the Full Blood Count (FBC), one of the most frequently ordered tests in primary care. They analyzed five specific metrics: Platelet count, MCV, MCH, MCHC, and RDW.

To handle data uncertainty and extract maximum value, they employed several novel feature engineering techniques:

  • Soft Range Definition: Recognizing that medical "normal" boundaries are somewhat arbitrary, they defined a "soft range" (a 2.5% buffer inside the standard normal limits) to capture borderline cases that might still indicate underlying pathology.
  • Temporal Labeling: They categorized results based on their proximity to the diagnosis date (e.g., within a 3-month window vs. earlier), creating six different versions of feature sets to test which temporal patterns were most predictive.
  • Combinatory Features: They created synthetic features by combining blood metrics (using additive and multiplicative factors) and incorporating biological sex to see if the interaction between different biomarkers improved the signal.
  • Machine Learning: They trained and validated three models—Decision Trees, AdaBoost, and XGBoost—using 10-fold cross-validation to predict 2-year survival.

Datasets

The study utilized a Victorian subset of the MedicineInsight dataset, a nationally representative primary care database in Australia.

  • Scale: The broader dataset includes 1.8 million patients and 85 million pathology results.
  • Study Cohort: The researchers identified a specific cohort of 472 lung cancer patients who had a recorded history for all five targeted hematological measures, ensuring no "missing data" assumptions were needed for the final model training.

Key Insights and Ideas

  • The Power of "Out-of-Range" (OOR): The mere presence of an abnormal blood result is a strong indicator. Patients with OOR results within three months of diagnosis had mortality rates up to 15% higher than those with normal results.
  • Gender-Specific Risks: The impact of abnormal results is not uniform across sexes. For example, an out-of-range platelet count in women was associated with a jump in mortality from 71.25% to 89.29%—a much sharper increase than that seen in men.
  • Borderline Matters: The "soft range" approach confirmed that values just inside the normal limit still carry predictive weight, suggesting that traditional binary "normal/abnormal" flags in lab reports may overlook early warning signs.
  • Robustness Amidst Scarcity: Even with a relatively small number of patients and only five blood metrics (and no data on cancer stage or treatment), the models were able to extract a consistent signal.

Key Results

  • Predictive Accuracy: The models achieved a classification accuracy of just under 65%. While this might seem modest in a vacuum, the authors argue it is a "strong indicator" given the high uncertainty of the data and the total lack of information regarding cancer stage or treatment type.
  • Feature Relevance: The study demonstrated that features derived from the timing of abnormal tests were more relevant than the raw values themselves.
  • Consistency: The generated features showed high consistency across different data subsets, proving that the methodology effectively handles the inherent uncertainty of primary care records.

Limitations and Future Work

The primary limitation is the sample size; 472 patients is relatively small for complex ML models, which likely capped the maximum achievable accuracy. Additionally, the dataset lacked clinical context such as the stage of the cancer at diagnosis or the specific treatments administered, both of which are massive determinants of survival.

Future directions include: * Applying the methodology to other cancer types (e.g., colorectal). * Incorporating more complex time-series analysis to look at the trajectory of blood results over years, rather than just snapshots. * Integrating additional primary care datasets (prescriptions, symptoms) to further reduce uncertainty and boost predictive power.

Summary updated

Abstract

More than 144 000 Australians were diagnosed with cancer in 2019. Diagnosing cancer in primary care is challenging due to the non-specific nature of cancer symptoms and its low prevalence. Understanding the epidemiology of cancer symptoms and patterns of presentation in patient's medical history from primary care data could be important to improve earlier detection and cancer outcomes. As past medical data about a patient can be incomplete, irregular or missing, this creates additional challenges when attempting to use the patient's history for any new diagnosis. Our research aims to investigate the opportunities in a patient's pathology history available to a GP, initially focused on the results within the frequently ordered full blood count to determine relevance to a future high-risk cancer prognosis, and treatment outcome. We investigated how past pathology test results can lead to deriving features that can be used to predict cancer outcomes, with emphasis on patients at risk of not surviving the cancer within 2-year period. This initial work focuses on patients with lung cancer, although the methodology can be applied to other types of cancer and other data within the medical record. Our findings indicate that even in cases of incomplete or obscure patient history, hematological measures can be useful in generating features relevant for predicting cancer risk and survival. The results strongly indicate to add the use of pathology test data for potential high-risk cancer diagnosis, and utilize additional pathology metrics or other primary care datasets even more for similar purposes.
BibTeX
@inproceedings{ristanoski2021handling,
  author = {Goce Ristanoski and Jon Emery and Javiera Martinez-Gutierrez and Damien McCarthy and Uwe Aickelin},
  title = {{Handling uncertainty using features from pathology: opportunities in primary care data for developing high risk cancer survival methods}},
  booktitle = {ACSW},
  year = {2021},
  url = {https://openreview.net/forum?id=qTAAKL0N3j},
}
Figure