Objective　To　propose　a　reconstruction　method　based　on　deep　learning　for　addressing　the　challenge　of　reconstructing　radial　artery　pulse　wave　from　fingertip　pulse　wave.　Methods　A　four-channel　data　acquisition　system　PowerLab　was　used　to　non-invasively　acquire　finger　pulse　waves　and　radial　artery　pulse　waves.　The　noise　source　in　the　pulse　wave　signals　were　analyzed,　and　the　stable　signal　waveforms　were　obtained　after　baseline　removal,　wavelet　transform　denoising,　and　normalization　preprocessing.　The　structure　and　parameters　of　the　variational　auto-encoder　(VAE)　network　model　were　designed.　The　model　was　trained　using　10-fold　cross-validation　on　data　from　744　subjects　to　establish　a　prediction　model　for　radial　artery　pulse　waves;　and　the　VAE　network　model　was　optimized　by　adjusting　hyperparameter　settings　of　learning　rate,　dropout,　and　regularization　term.　Results　The　results　from　the　reconstruction　and　synchronous　detection　of　radial　artery　pulse　waves　in　186　subjects　showed　that　for　reconstructing　radial　artery　pulse　waves　from　low-　and　high-resistance　fingertip　pulse　waves,　the　5%　K　difference,　20%　K　difference,　total　variance　of　K　difference,　and　FIT　were　49.10%,　96.70%,　89.74,　and　75.80%　when　using　VAE　network　model,　and　those　were　48.50%,　94.50%,　73.74,　and　66.30%　when　using　VAE　optimization　model.　Conclusion　The　VAE　network　model　and　its　optimization　method　can　be　used　for　radial　artery　pulse　wave　reconstruction,　with　high　reconstruction　accuracy,　strong　robustness　and　generalization　ability.　©　2024　Editorial　Department　of　Chinese　Journal　of　Medical　Physics.　All　rights　reserved.