Clardia - Prediction of Diabetes Using Non Invasive Photoplethysmography (PPG) Measurements & Physiological Characteristics.


Photoplethysmography (PPG) can be identified as a non invasive, inexpensive, optic technique which measures the blood volume changes in blood vessels through which oxygen saturation, blood pressure, cardiac output could be measured. In recent researches it has been identified that PPG is a promising technique towards early screening of diseases as the PPG waveform possess significant information embedded within.

Diabetes is a result of abnormally high levels of sugar in the blood, which results in many long term complications. The early detection of Diabetes is of utmost importance to reduce risks and proper health management. The research focuses on analyzing extracted PPG signals of users, their physiological & demographic characteristics using Machine Learning Techniques, in order to predict Diabetes.

The regular measurement of health parameters is important towards identification and diagnosis of such diseases. The rapid development of smart wearable devices and mobile phones embedded with PPG sensors has enabled users to easily obtain the required health measurements towards early detection of diseases.

At present the PPG sensors are mainly been used for Heart Rate estimation, whereas non invasive BP, Blood Glucose estimation and disease prediction are comparatively novel fields of research. The latest clinical studies have been able to detect diabetes with an accuracy of 85%.(Ballinger et al 2018) The successful research in the area would ensure great value addition in the fields of modern healthcare & technology.

Aim & Objectives

The aim of this research is to develop a machine learning model capable of detecting diabetes using PPG signal data and other physiological & demographic characteristics age, gender, height and weight.


  • Identify appropriate features for the development of the model. (PPG, Physiological & demographic)
  • Identify suitable machine learning techniques in order to develop a model for the prediction of diabetes.
  • Validate the models using test data.


  1. Review existing literature related to biomedical signal processing and machine learning approaches towards diabetes prediction to identify possible features.
  2. Develop models to predict diabetes using the features identified in step 1.
  3. Develop models towards automatic feature learning for diabetes prediction.
  4. Implement the model using the identified methodology and bring the quality and accuracy of the predictions to a level where the model can be applied practically.
  5. Suggest methods where the research can be applied practically with the use of general wearable devices embedded with PPG sensors.

The research would be carried out using the dataset, ‚ÄčA new, short-recorded photoplethysmogram dataset for blood pressure monitoring in China (‚ÄčLiang, Yongbo; Liu, Guiyong; Chen, Zhencheng; Elgendi, Mohamed (2018) published in the Nature Journal under a creative common license.


Please find the final thesis of the above project here for the results and discussion. The code base of the project here.