HANSE POST
The HANSE-POST study builds upon the HANSE LCS trial, a holistic northern German interdisciplinary lung cancer screening study. The interdisciplinary HANSE-POST network consists of 4 subprojects: (SP1) Conventional LDCT adjusted LC risk prediction models (SP2) Implementation and validation of risk adapted screening intervals (SP3) Socioeconomic and implementation modelling (SP4) Machine Learning-based individualised screening intervals.
The overall aim of the network is to show that individualised post-LDCT risk-based LCS intervals are more effective compared to annual screening. Therefore, we conduct the prospective randomised HANSE-POST trial to answer the question if risk-based LCS intervals have indeed an equal proportion of advanced stage lung cancers compared to annual screening (SP2). For choosing the best suited risk model we investigate two approaches: (1) a conventional modelling approach based on prior work (Robbins 2019 PMID: 30976808; Robins 2022 PMID: 34648946) with refinement and recalibration to the HANSE cohort (SP1) and (2) a Deep Learning approach, which is based on promising recent retrospective work from the US (Mikhael 2023 PMID: 36634294, SP4). After comparison of the two approaches the winning model will be prospectively applied to define the low and high risk groups in the prospective HANSE-POST trial. Furthermore in SP4, novel Machine Learning models based on Radiomics and on Deep Learning with an anomaly detection approach will be explored using LDCT images, lung function and blood pressure as well as extensive questionnaire data to predict lung cancer risk, based on the available NLST, HANSE and LUSI datasets used for training, testing and external validation.
Finally, we will employ explainable AI (XAI) methods to determine how radiomics and Deep Learning based classifiers define the phenotype of low risk and high risk on the best performing models in SP4.
Moreover, we will use microsimulation methods to model the effects of different screening scenarios on a population level. The risk prediction models developed in SP1 and SP4 will be incorporated in the simulation model to schedule individual screening intervals for the LCS eligible synthetic model population. In 500+ simulated scenarios, the lifetime health outcomes and costs of ten years of LCS will be computed. Primary outcomes of the cost-effectiveness analyses will be additional costs per quality-adjusted life year (QALY) gained, per life year (LY) gained or per preventable death.
Improve PLCO LC risk prediction model using LDCT data:
For comparison with related work, an extended version of the PLCO risk prediction will be developed using NLST data. In following steps, data from the German LUSI trial and from the first two screening rounds in HANSE will be used for independent validation, and for optimising (re-)calibration specifically to the participants of the HANSE trial.
Identify most suitable Machine Learning-based risk estimation approach:
Two different Machine Learning approaches (one relying on LDCT, the other additionally includes radiomic and features like questionnaires) will be trained with data from the HANSE study as well as the LUSI dataset provided by the DKFZ. To evaluate the generalizability of the trained models, external validation using NLST data is crucial. Vice versa, we plan to train the same Machine Learning models with NLST data and evaluate the performance in risk prediction of HANSE and LUSI data. Furthermore, we will compare our approach with related work like Sybil (Mikhael 2023 PMID: 36634294). Since Sybil is trained with NLST data, we need to use the same training data to ensure a fair comparison.
Comparing conventional LC risk prediction models with Machine Learning approaches:
A benchmarking between conventional risk models relying on participant data (e.g. smoking history) and conventional risk models incorporating LDCT data (e.g. emphysema score) on the one site as well as several Machine Learning based risk estimation models is planned. This comparison will include multiple datasets (HANSE, LUSI, NLST) to identify the best risk estimation model. The best model is defined by highest sensitivity and specificity.
Explaining Machine Learning –based lung cancer risk estimation:
It is expected, that a Machine Learning approach outperform conventional risk prediction models. However, in contrast to conventional risk models, especially Deep Neural Networks can be referred to as a black box. To enhance the trust in such AI models and thus foster the translation into a national LCS program, explainable AI methods will be used to reveal the path from input data to risk estimation.
Starting from a collection of Machine Learning models trained only with HANSE, LUSI and NLST as well as models trained with combinations of these datasets, we aim to identify the variables and image regions actually being crucial for LC risk prediction. In a suitable model, the identified variables and image regions should be the same in HANSE, LUSI and NLST.
Estimate and compare population health outcomes, cost-effectiveness (CE) and budget impact (BI) of LCS strategies
The objective is the identification of optimal thresholds for risk-adjusted screening intervals. Population-based health outcomes and costs of 500+ LCS scenarios with varying risk factor thresholds (VTD, nodule volume, background risk score etc.) for risk-adjusted intervals as well as effects of no LCS, annual and biennial LCS will be simulated. In each scenario, the lifetime health outcomes and costs of ten years of LCS will be computed.
Jens Vogel-Claussen - Hannover Medical School (Coordinator)
Hinrich Winther- Hannover Medical School
Maximilian Zubke - Hannover Medical School
Sabine Bohnet - University Hospital Schleswig-Holstein
Rudolf Kaaks - German Cancer research Center
Alexander Kuhlmann - Martin Luther University Halle-Wittenberg
Martik Reck - Lung Clinic Grosshansdorf