Can AI Assist Public Health Case Investigators? The Secret of Epira, the Rule-Aware AI

An illustrative representation of an AI interview system talking with a patient and collecting data in a public health environment
AI Summary

Exploring the technical features and evaluation results of Epira, a rule-controlled LLM interview system developed to accurately assist with the arduous tasks of case investigation and contact tracing in public health.

Imagine this. One day, you suddenly test positive for an infectious disease. The public health center contacts you and asks, “Who did you meet over the past week? Where did you visit?” You frantically search your memory in a panic, breaking out in a cold sweat because exact times and locations refuse to surface.

On the front lines of preventing infectious disease spread, case investigation and contact tracing are core pillars of public health, but they are also processes requiring immense effort and labor [Source 1, Source 2]. This is because public health professionals must individually converse with patients to extract information without omissions.

Can artificial intelligence (AI) assist with this arduous process? A system called Epira, recently developed by researchers, provides a clear answer to this question [Source 1, Source 2]. Today at MindTickleBytes, we examine how this intriguing AI interview system works and why it is gaining attention in the medical field.


Why Is This Important?

During every infectious disease outbreak, public health centers and quarantine authorities interview patients and trace contacts day and night. However, conventional methods have clear limitations.

First, web questionnaire methods are too rigid. When respondents do not understand a question or find it ambiguous and need additional explanation, web forms cannot respond flexibly. Consequently, frustrated respondents often abandon the survey halfway or input incomplete information [Source 1, Source 2].

Second, introducing generic open-ended AI chatbots that converse freely like humans creates another problem. There is a risk that the AI might get carried away, omit quarantine protocols (such as mandatory question lists and safety guidelines), or confuse patients with erratic answers [Source 1, Source 2].

Simply put, the quarantine field urgently needed a tool that possesses both “the ability to communicate smoothly as if talking to a person, while strictly never missing predetermined medical protocols and mandatory items” [Source 1, Source 2]. Epira was created precisely to achieve both of these goals [Source 1, Source 2].


Easy to Understand: How Does Epira Work?

To use an analogy, Epira is like “a veteran investigator holding top-tier guidelines in hand.”

General Large Language Models (LLMs) converse very flexibly, but occasionally say erratic things or violate rules. Epira combines this AI’s exceptional ‘conversational ability’ with brakes and navigation in the form of a ‘rule-based controller’ [Source 1].

[Respondent (Patient)] 
      ↕ (Natural conversation)
[LLM Interviewer (Qwen2.5 or GPT-5.4)]
      ↕ (Enforce mandatory questions & safety protocols)
[Rule-based Controller] → [Structured Patient Record Data]

Simply put, if a patient says, “I don’t remember well who I met yesterday,” instead of stiffly throwing an error, the AI smoothly guides the conversation by saying, “It might help to recall the name of the restaurant you visited last evening or your credit card transaction history.” Behind the scenes, however, the rule-based controller operates to strictly ensure that “required questions that must legally or quarantine-wise be asked, safety handoffs, and structured case records” are filled out without a single omission [Source 1].

Furthermore, even if a novel infectious disease emerges, Epira can immediately perform customized interviews tailored to that disease simply by uploading a new questionnaire file without needing to modify complex code [Source 1].


Research and Experimental Results: What Did It Prove?

To validate Epira’s performance, the researchers conducted two rigorous studies [Source 1]. This research was conducted by researchers including Ishita Asnani, Shaghayegh Fazliani, Zachary Ryan, Kriyoshig Prasai, Chirag Kumar, Indra Gonzalez Ojeda, Marc Lipsitch, and Madeleine Udell [Source 7].

Study 1: Simulation-Based Large-Scale Session Analysis

The researchers created 108 simulated cases containing predefined ground-truth data for two diseases, measles and gonorrhea. They divided these into Epira conditions and standard open-ended conditions, conducting a staggering 864 interview sessions [Source 1].

The results were remarkable:

  • The Epira system using a local language model as its backend accurately captured 88.9% of required essential information fields and collected 60.1% of actual contact information [Source 1].
  • Conversely, ‘open-ended interviewers’ that were allowed to converse freely without rule control using the exact same model recovered 31.1 to 41.9 percentage points fewer essential information fields compared to Epira [Source 1].

This experiment clearly demonstrated that ‘protocol control’ is the core factor determining data completeness, surpassing the inherent intelligence of the AI model itself [Source 1].

Study 2: Usability Evaluation with Real Volunteers

In the second study, 72 volunteers were recruited to conduct actual interviews, evaluating completion rate, usability, acceptability, and trust [Source 1]:

  • The interview completion rate recorded a remarkable 100% [Source 1].
  • The System Usability Scale received a high score of an average of 82.95 $\pm$ 16.17 out of 100 [Source 1].
  • User trust also gained a positive evaluation, recording an average of 3.85 $\pm$ 0.77 out of 5 [Source 1].

Notably, the researchers stated that the final records generated by open-source models running in local environments and cloud-based large models exhibited similar high levels of quality [Source 1]. This opens up the possibility of securely processing sensitive patient medical data in privacy-preserving local environments without sending it to external clouds [Source 1].


Current Status and Limitations: What We Need to Know

Epira is an exemplary case showing how artificial intelligence can create tangible value in the public health sector. It is receiving attention from academia and the field for simultaneously overcoming the limitations of web surveys and the risks of open-ended chatbots [Source 1, Source 2].

However, just because the technology is advanced does not mean human touch will completely disappear from medical settings immediately. The researchers emphasize that while records collected in local deployment environments significantly help protect privacy, the finally routed information must necessarily undergo human review (by public health professionals) to be safe [Source 1]. While AI performs the role of an excellent drafter and diligent assistant, a collaborative structure where humans hold final judgment and responsibility is essential.


What Lies Ahead?

When a novel infectious disease or public health crisis strikes in the near future, public health authorities may no longer need numerous personnel tied to telephones repeatedly conducting basic questionnaires one by one. A world is approaching where, the moment new disease guidelines are released, they can be uploaded to the system to deploy custom AI investigators in seconds, and patients can comfortably and safely report their symptoms and contacts through conversation with AI anytime, anywhere [Source 1].

Technology may seem cold and complex, but when it meets the right rules and control mechanisms, it becomes a powerful shield that warmly protects our daily lives. Epira’s successful evaluation stands as yet another milestone showing that AI is evolving into a reliable companion safeguarding human health and safety [Source 1, Source 2].


References

  1. Ishita Asnani, Shaghayegh Fazliani, Zachary Ryan, Kriyoshig Prasai, Chirag Kumar, Indra Gonzalez Ojeda, Marc Lipsitch, Madeleine Udell, Design and Evaluation of Epira: An Automated Interview System for Case Investigation and Contact Tracing, arXiv:2611.25577 [Source 1, Source 7]
  2. Design and Evaluation of Epira: An Automated Interview… [Source 2]
  3. Design and Evaluation of Epira: An Automated Interview System … [Source 5]
  4. Machine Learning - arXiv.org [Source 7]
AD
Test Your Understanding
Q1. What is the most core feature that distinguishes Epira from conventional open-ended chatbots or web surveys?
  • Works only without an internet connection
  • Combines LLM conversational capability with a rule-based controller to enforce mandatory questions and safety protocols
  • Supports voice calls with patients only
Epira combines the flexible conversational ability of an LLM with a rule-controlled controller to prevent missing mandatory questions and strictly adhere to safety protocols [Source 1].
Q2. Which of the following is correct regarding the backend model configurations tested in the research?
  • Only the cloud model (GPT-5.4) was tested
  • Only the local model (Qwen2.5) was tested
  • Both local (Qwen2.5) and cloud (GPT-5.4) backends were tested
The Epira study tested and compared both local model (Qwen2.5) and cloud model (GPT-5.4) backends [Source 1].
Q3. What was the interview completion rate evaluated with 72 volunteers in Study 2?
  • 100%
  • 88.9%
  • 60.1%
In Study 2, the interview completion rate recorded 100% [Source 1].
Can AI Assist Public Health...
0:00