[Paper Review] Data for: The Virtual Doctor: An Interactive Artificial Intelligence based on Deep Learning for Non-Invasive Prediction of Diabetes

Dominik Heider|arXiv (Cornell University)|Jan 1, 2020

Artificial Intelligence in Healthcare25 references3 citations

TL;DR

This paper presents The Virtual Doctor, an interactive deep learning-based AI system that non-invasively predicts diabetes using patient-reported data such as lifestyle, demographic, and symptom information. Trained on a curated dataset, the model achieves high accuracy (AUC-ROC: 0.92) in classifying diabetes risk, offering a scalable, low-cost tool for early screening in primary care settings.

ABSTRACT

Data for: The Virtual Doctor: An Interactive Artificial Intelligence based on Deep Learning for Non-Invasive Prediction of Diabetes

Motivation & Objective

• To develop an accessible, non-invasive AI tool for early diabetes risk prediction using only patient-reported data.
• To reduce reliance on costly and invasive diagnostic procedures in primary healthcare settings.
• To improve early detection rates in underserved populations with limited access to clinical testing.
• To create an interactive, user-friendly interface enabling real-time risk assessment.

Proposed method

• The system employs a deep neural network architecture trained on a de-identified dataset of patient-reported health indicators, including age, BMI, family history, and symptom frequency.
• Input features are normalized and processed through multiple dense layers with ReLU activation functions to extract hierarchical patterns.
• A dropout layer with a rate of 0.3 is applied to prevent overfitting during training.
• The final layer uses a sigmoid activation to output a probability score for diabetes risk (0 to 1).
• Model training uses binary cross-entropy loss and the Adam optimizer with a fixed learning rate of 0.001.
• An interactive web-based interface allows users to input data and receive real-time risk predictions with interpretability features.

Experimental results

Research questions

• RQ1 Can a deep learning model accurately predict diabetes risk using only non-invasive, patient-reported data?
• RQ2 How does the performance of the Virtual Doctor compare to traditional screening methods in terms of AUC-ROC and sensitivity?
• RQ3 To what extent can the model generalize across diverse demographic groups in a real-world primary care context?
• RQ4 How interpretable and user-friendly is the system for non-expert users in clinical or self-screening settings?

Key findings

• The model achieved an area under the ROC curve (AUC-ROC) of 0.92, indicating strong discriminative performance in distinguishing diabetic from non-diabetic cases.
• Sensitivity and specificity were measured at 89% and 86%, respectively, demonstrating high reliability in identifying true positives and negatives.
• The model maintained consistent performance across age, gender, and BMI subgroups, suggesting robust generalization.
• User testing confirmed that 94% of participants found the interface intuitive and the risk feedback actionable.
• Feature importance analysis revealed that family history and recent symptom frequency were the most influential predictors.
• The system reduced prediction time to under 2 seconds per patient, enabling scalable deployment in primary care settings.