Graduation Year


Document Type




Degree Name

MS in Computer Science (M.S.C.S.)

Degree Granting Department

Engineering Computer Science

Major Professor

Dmitry Goldgof, Ph.D.

Committee Member

Yu Sun, Ph.D.

Committee Member

Mark Last, Ph.D.


Detection, Frequency Modulation, Medical Data, Prediction, Convolutional Neural Networks


Spectrograms extract frequency components from a signal. Spectrograms have beenin use for a long time mainly to analyze frequency components in audio signals. Typically, these audio signals have a very high sampling rate, various frequency components and high frequency variability with time. Vital signs on other hand have very low sampling rate with no frequency variability. This work explores if spectrograms can be used to analyze and recognize patterns from vital signs signals.

As mentioned above, spectrograms deal with frequencies. More the variability of frequency, better the patterns emerge when spectrograms are applied on the signals. As vital signs lack the frequency variability, this work uses frequency modulation to introduce frequency variability. The amplitude of vital signs signal is encoded as frequency of another signal on which spectrogram is applied to understand the frequency variability which is directly proportional to the amplitude of the vital signs signal.

Then, the generated spectrograms are given as an input to convolutional neural networks to extract features and classify the spectrograms. The efficacy of this method is tested on 4 different datasets for prediction and detection. The proposed method was able to achieve an accuracy of 91.55% and an AUC of 0.92 for prediction task and had better overall precision, recall, F1-score compared to the baseline method. It also achieved an accuracy of 91.67 and AUC of 0.92 on classification or detection task and also it performed better than baseline method in all the metrics such as precision, recall, F1-score for detection.