Research on Accurate Prediction Model for Cardiovascular Disease Mortality Based on PCA-Combined Grey Neural Network

Abstract Background Cardiovascular diseases (CVD) are a leading global chronic threat to human health, marked by high mortality, prolonged courses, and complex pathogenesis. Data from the 2024 Report on Cardiovascular Diseases in China shows CVD-related deaths account for over 40% of total deaths in the country, with an annual upward trend. Meteorological factors (temperature, humidity, air pressure, wind speed) and air pollutants (PM2.5, PM10, SO₂, NO₂) are key triggers, inducing physiological changes like vasoconstriction to raise acute CVD risks. While advanced monitoring provides massive data for prediction, traditional models struggle with multi-factor nonlinear coupling, data redundancy, and poor adaptability to mortality fluctuations, failing to meet clinical and public health needs. Purpose This study aimed to build an optimized model to boost CVD mortality prediction accuracy. Methods The study used data from Shenzhen (Jan 1, 2023–Feb 5, 2024): meteorological data (4 indicators), air pollutant data (6 indicators from 11 stations), and CVD mortality data (daily counts via ICD-10 codes), forming 400 valid 10-input-1-output datasets. Preprocessing included: cumulative lag analysis (τ=1, strongest correlation), min-max normalization ([0,1] mapping), and 8:2 training-test (320:80) split. A three-layer PCA-combined grey neural network was built: preprocessing layer used KMO (0.762) and Bartlett’s (Sig.0.000) tests, extracting 3 principal components (82.32% cumulative contribution); core layer had PCA-SGNN (GM(1,1) trend + BP residual correction, 10-window GM(1,1), 4-6-1 BP) and PCA-PGNN (GM(1,1)+BP parallel prediction, 3-7-1 BP, MAPE-based weighting); fusion layer did dynamic weight allocation (by sub-module accuracy) and error feedback correction (1.2×MAE threshold, weight adjustment). Results The optimized model performed excellently: MAPE 11.86%, prediction accuracy (≤10% error) 91.78%, acceptability (≤15% error) 94.52%. It outperformed single models (BP: ∼25.32% MAPE, 68.15% accuracy; GM(1,1): 22.18% MAPE, 72.43% accuracy) and basic combined models (PCA-BP: 18.65% MAPE, 80.27% accuracy; SGNN:16.34% MAPE,83.51% accuracy; PGNN:14.21% MAPE,86.79% accuracy). It had a 1.23 error standard deviation (lower than others) and strong fluctuation adaptability (≤1.5 error in peaks/troughs), with posterior error test (C=0.32, P=0.96) confirming robustness. Conclusions The optimized model overcomes traditional flaws: PCA reduces dimensionality while retaining 82.32% information; dual sub-modules complement trend and fluctuation capture; fusion refines results. It provides reliable support for CVD risk early warning, medical resource allocation, and public health decisions. Future optimization with personal factors and cross-city validation will enhance its applicability.