Abstract
Many COVID-19 patients developed prolonged symptoms after the infection, including fatigue, delirium, and headache. The long-term health impact of these conditions is still not clear. It is necessary to develop a way to follow up with these patients for monitoring their health status to support timely intervention and treatment. In the lack of sufficient human resources to follow up with patients, we propose a novel smart chatbot solution backed with machine learning to collect information (i.e., generating digital diary) in a personalized manner. In this article, we describe the design framework and components of our prototype.
Introduction
The COVID-19 pandemic fundamentally shifted how patients seek medical care, with many opting to stay home to avoid infection risks. This created a dangerous "information gap": healthcare providers often remain unaware of a patient's deteriorating condition until it becomes an emergency. While telemedicine and surveys exist, they face significant "bandwidth" issues—nurses can only see a limited number of patients per day, and static surveys often fail to capture the nuance of evolving symptoms.
To address this, we developed a smart chatbot prototype that bridges the gap between self-isolating patients and overstretched healthcare providers. By combining a literature-derived knowledge graph with a conversational NLP pipeline, the system creates personalized "digital diaries" to track symptom progression and provide evidence-based information to users.
System Architecture and Methods
The proposed framework integrates literature mining with real-time patient interaction through a multi-layered technical stack designed to process natural language and build a structured knowledge base.
The system uses a standard preprocessing flow (tokenization, lemmatization, and stop-word removal) followed by Named Entity Recognition (NER). Rather than training from scratch, the team utilized pre-trained medical models: Med7 for clinical labels (drug, dosage, etc.) and scispaCy for broader biomedical entities like diseases and chemicals. The conversational agent itself is built on the Rasa NLU framework, which handles intent classification and entity extraction from user input.
Entities extracted from scientific literature, specifically the CORD-19 dataset containing over 250,000 articles, form the nodes of a global Knowledge Graph (KG). Edges represent relationships, weighted by the frequency of co-occurrence in the literature or semantic links derived from part-of-speech tagging. This allows the bot to "learn" about symptoms—such as the loss of taste or smell—from literature before its first patient interaction.
Personalized Health Trajectories
A core innovation of this framework is the generation of personalized "subgraphs" for every patient. This maps an individual's reported symptoms onto the larger literature-based KG to visualize their specific disease trajectory over time.
By treating a patient's history as a subgraph, the system can eventually "mine" these trajectories. If a patient's early symptoms match a known pattern in the database, the system could potentially predict future symptoms (e.g., predicting the onset of specific lab result changes). This turns a casual chat into a structured "digital diary"—a clinical-grade history that a doctor can review in seconds, saving significant time during actual consultations.
Prototype Evaluation and Results
The paper presents a functional prototype demonstrating key qualitative results:
- Successful Entity Mapping: The system demonstrated the ability to take natural language input (e.g., "My chest hurts") and correctly record it as a clinical entity in the patient's profile.
- Contextual Follow-ups: The bot successfully used historical data to ask personalized questions, such as: "You mentioned in our last conversation that you have symptoms of nose, coughing, and headache. How do you feel about them now?"
- Evidence-Based Redirection: Due to legal constraints, the bot does not prescribe medicine. Instead, it successfully demonstrated redirecting users to specific, relevant sentences from scientific literature and official guidelines.
Limitations and Future Work
While promising, the prototype faces several hurdles. Legal and safety constraints prevent the chatbot from providing direct medical advice or prescriptions. Additionally, to truly predict patient trajectories, the system requires a much larger cohort of patient data than was available during the prototype phase. Future work will focus on improving generalization to other time-sensitive conditions and implementing a seamless "hand-off" mechanism to alert human doctors in real-time if a patient's trajectory looks life-threatening.
Summary updated