Early detection of prostate gland and breast cancer risk based on routine check-up data using survival analysis trees for left-truncated and right-censored data

Dan Coster, Eyal Fisher, Shani Shenhar‐Tsarfaty, Tehillah S. Menes, Shlomo Berliner, Ori Rogowski, David Zeltser, Itzhak Shapira, Eran Halperin, Saharon Rosset, Malka Gorfine, Ron ShamirmedRxiv 2021 · DOI

Abstract

Objective: To predict breast cancer (BC) and prostate gland cancer (PGC) risk among healthy individuals by analyzing routine laboratory measurements, vital signs, and age. Materials and Methods: We analyzed electronic medical records of 20,317 healthy individuals who underwent routine checkups, encompassing more than 600 parameters per visit, and identified those who later developed cancer. We developed a novel ensemble method for risk prediction of multivariate time series data using a random forest model of survival trees for left-truncated and right-censored data. Results: Using cross-validation, our method predicted future PGC and BC 6 months before diagnosis, achieving an area under the ROC curve of 0.62±0.05 and 0.6±0.03 respectively, better than standard random forest, Cox-regression model, and a single survival tree. Our method can complement existing screening tests such as clinical breast examination and mammography for BC, and help in detection of subjects that were missed by these tests. Discussion: Computational analysis of results of routine checkups of healthy individuals can improve the detection of those at risk of cancer development. Conclusion: Our method may assist in early detection of breast and prostate gland cancer.

Introduction

Early detection is the "holy grail" of oncology, significantly improving prognosis and survival rates. However, current screening strategies—like mammography for breast cancer (BC) or PSA testing for prostate gland cancer (PGC)—can be costly, labor-intensive, and suffer from variable accuracy. While Electronic Medical Records (EMR) contain a wealth of routine data (blood counts, metabolic panels, vitals), this data is difficult to model because it is multivariate, longitudinal, and subject to left-truncation and right-censoring. This paper introduces TVsuRF, a novel way to harness this "messy" routine data to create a personalized risk score that complements existing clinical tests.

Methods

The researchers developed TVsuRF (Time-Varying SUrvival Random Forest) to address the complexities of longitudinal medical data.

The "Pseudo-Object" Concept

To handle time-varying covariates, the researchers split each subject’s history into disjoint intervals. Each interval is treated as a "pseudo-object" with its own entry time (left-truncation) and exit time (either a cancer event or censoring). Unlike standard decision trees that use Gini impurity, TVsuRF uses conditional inference trees with statistical hypothesis testing (Pan's permutations) to select optimal splitting variables, allowing for robust comparisons of survival curves.

Datasets and Features

The study utilized the TAMCIS (Tel-Aviv Medical Center Inflammation Survey) dataset, containing records for 20,271 healthy adults across 50,497 visits. Cancer outcomes were verified via the Israeli National Cancer Registry (INCR). The models utilized 20 covariates for BC (primarily Complete Blood Count parameters) and 28 for PGC (adding metabolic panels, lipids, and vitals).

Table 1: Baseline characteristics and clinical parameters for the Breast Cancer (BC) cohort.
Table 2: Baseline characteristics and clinical parameters for the Prostate Gland Cancer (PGC) cohort.

Key Results

TVsuRF consistently outperformed traditional Cox regression and standard Random Forest models, particularly in shorter-term prediction windows.

  • Predictive Power: The model successfully predicted cancer risk 6 months (183 days) before diagnosis, achieving an Area Under the ROC Curve (AUROC) of 0.62 ± 0.05 for PGC and 0.60 ± 0.03 for BC.
  • Top Predictors:
    • BC: Mean Corpuscular Volume (MCV), Monocytes, Mean Platelet Volume (MPV), and Age.
  • PGC: Alkaline Phosphatase (ALP), LDL cholesterol, Age, Calcium, and Glucose.
  • Filling Screening Gaps: In subset analyses, the model identified high-risk women who had tested "normal" in Clinical Breast Exams (CBE) and mammography, suggesting TVsuRF could act as a vital safety net in clinical settings.

Discussion and Conclusion

The success of the model using Complete Blood Count (CBC) and metabolic data suggests that even early-stage cancer causes systemic changes detectable in routine labs. Variables like MCV and ALP emerged as high-importance features, despite being standard parts of a routine check-up rarely used for cancer risk stratification.

While limited by class imbalance and varying data density, TVsuRF demonstrates that computational analysis of routine checkups can improve the detection of those at risk. This method may significantly assist in the early detection of breast and prostate gland cancer by providing a personalized, data-driven "second opinion" to standard screening protocols.

Summary updated

Abstract

ABSTRACT Objective To predict breast cancer (BC) and prostate gland cancer (PGC) risk among healthy individuals by analyzing routine laboratory measurements, vital signs and age. Materials and Methods We analyzed electronic medical records of 20,317 healthy individuals who underwent routine checkups, encompassing more than 600 parameters per visit, and identified those who later developed cancer. We developed a novel ensemble method for risk prediction of multivariate time series data using a random forest model of survival trees for left truncated and right-censored data. Results Using cross-validation, our method predicted future PGC and BC 6 months before diagnosis, achieving an area under the ROC curve of 0.62±0.05 and 0.6±0.03 respectively, better than standard random forest, Cox-regression model and a single survival tree. Our method can complement existing screening tests such as clinical breast examination and mammography for BC, and help in detection of subjects that were missed by these tests. Discussion Computational analysis of results of routine checkups of healthy individuals can improve the detection of those at risk of cancer development. Conclusion Our method may assist in early detection of breast and prostate gland cancer.
BibTeX
@misc{coster2021early,
  author = {Dan Coster and Eyal Fisher and Shani Shenhar‐Tsarfaty and Tehillah S. Menes and Shlomo Berliner and Ori Rogowski and David Zeltser and Itzhak Shapira and Eran Halperin and Saharon Rosset and Malka Gorfine and Ron Shamir},
  title = {{Early detection of prostate gland and breast cancer risk based on routine check-up data using survival analysis trees for left-truncated and right-censored data}},
  year = {2021},
  doi = {10.1101/2021.01.11.21249491},
  howpublished = {medRxiv},
}
Figure