Public Health Need Understanding where lung cancer rates are rising, falling, or remaining stable is essential for allocating cancer prevention resources, targeting smoking cessation programs, and planning screening infrastructure. County-level incidence data provides far more actionable geographic resolution than state or national averages.
Study Scope Guo and colleagues used lung cancer incidence data from 1,013 US counties spanning 2000 to 2021 from the SEER (Surveillance, Epidemiology, and End Results) database to compare six different prediction models, including both established machine learning methods and a novel physics-inspired technique called Dynamic Mode Decomposition (DMD).
Why Compare Methods No single model is universally best for time-series prediction. Different algorithms have different strengths - some are better at capturing linear trends, others at modeling spatial correlations, and others at handling complex nonlinear dynamics. A head-to-head comparison on real-world cancer data provides guidance for public health agencies choosing forecasting tools.
What DMD Does Dynamic Mode Decomposition is a technique originally developed in fluid dynamics. It decomposes a temporal dataset into spatial modes (patterns) and their corresponding temporal dynamics (oscillation frequencies and growth/decay rates). Applied to cancer incidence data, DMD identifies coherent spatial-temporal patterns of how incidence changes across counties over time.
Competing Models The six models compared were: DMD, Time Series (ETS - exponential smoothing state space model), Random Forest (RF), Gradient Boosting Machine (GBM), Support Vector Machine (SVM), and Linear Regression. This covers classical statistical forecasting, tree-based ensemble methods, kernel methods, and the novel DMD approach.
Evaluation Metrics Models were evaluated using Root Mean Square Error (RMSE) - measuring average prediction error magnitude - and Spearman correlation - measuring how well the model preserves the rank ordering of incidence rates across counties. Both metrics capture different aspects of forecasting quality.
RMSE Rankings Time Series (ETS) achieved the lowest RMSE at 15.3, indicating the smallest average prediction error in absolute incidence rates. DMD and Random Forest tied with RMSE of 20.5, while GBM, SVM, and Linear Regression performed worse. For raw incidence rate prediction accuracy, statistical time series methods thus outperformed DMD.
Spearman Correlation DMD achieved the highest Spearman correlation of 0.706, meaning it best preserved the relative ranking of counties by incidence. This suggests DMD is particularly good at identifying which counties will have relatively high versus low incidence, even if its absolute number predictions are less precise than ETS.
Geographic Pattern Consistent with known epidemiological data, Kentucky consistently showed the highest lung cancer incidence (driven by high smoking rates and occupational exposures), while California, New Mexico, Utah, and Idaho had the lowest rates. Models that captured these geographic patterns had better Spearman correlations.
Interpretable Spatial Modes Unlike black-box machine learning models, DMD produces interpretable outputs - specific spatial modes that reveal coherent geographic patterns in how lung cancer incidence evolves. Public health officials can examine these modes to understand which regional clusters move together, potentially reflecting shared risk factor trends.
Computational Efficiency DMD requires significantly less computational resources than training deep learning or gradient boosting models, making it practical for agencies with limited computational infrastructure that need to regularly update forecasts as new data becomes available.
Linear Approximation Limitation DMD's key weakness is that it models the underlying system as linear. When cancer incidence dynamics are driven by sudden nonlinear changes - such as the introduction of new screening guidelines, major policy changes, or rapid shifts in smoking prevalence - DMD's linear approximation may fail to capture abrupt transitions.
Resource Allocation County-level incidence predictions can guide where to place mobile CT screening units, where to concentrate smoking cessation outreach, and where to prioritize radon testing programs - all of which directly reduce lung cancer mortality in high-risk communities.
Model Selection Guidance This study provides practical guidance for public health agencies: use ETS or similar time series methods when precise absolute incidence numbers matter most, and use DMD when geographic pattern analysis and identifying co-varying regional clusters is the primary goal.
Hybrid Future Directions Combining DMD's interpretable spatial modes with the predictive precision of machine learning models (for example, using DMD-identified spatial patterns as features in a subsequent ML model) could yield both accurate predictions and interpretable geographic insights - an area the authors identify for future research.