In this paper, we introduce a neural network (NN)-based framework aimed at classifying sensing and communication signals at base stations, improving the efficiency of integrated sensing and communication (ISAC) systems in a bistatic configuration. The framework leverages a key mathematical insight: the Hankelized matrix formed from an equidistantly sampled signal of sparsely superimposed radio waves exhibits a low-rank property, whereas a frequency-modulated signal lacks this characteristic. It ensures that, even in practical environments, the Hankelized matrix of a sensing or communication channel statistically retains the relevant information. Hence, we use the singular values of the Hankelized matrix as the input to the neural NN, while the output is a one-hot encoded vector indicating whether the received signal is intended for sensing or communication. We investigate three scenarios where the communication and sensing signals either use the same or different waveforms in terms of the detection performance of the communication signals. The results demonstrate that the proposed method outperforms existing approaches in classification performance across all scenarios, regardless of whether the communication and sensing signals utilize the same waveform or not. The framework achieves a detection rate of over 95% even at an SNR of 0 dB. Notably, the network performs well in terms of a small number of pilot symbols, a small number of training dataset, and dynamic environments.