Moving towards clinically validated Neuro-Symbolic AI

Generative AI has created a new paradigm in AI development, and even in healthcare—which typically evolves slowly—we are witnessing exciting advancements almost daily. These innovations inspired us to rethink how we approach our product strategy at Infermedica.

However, unlike other, non-regulated areas of our lives, applying Generative AI in healthcare profoundly impacts patients; it can either help or harm. Thus, it requires rigorous safety considerations, clinical validation, supervision, and regulatory approval before we can safely "take our hands off the steering wheel." After all, I wouldn't want my children to be examined by someone who’s a "magician" if there’s even a slight risk the magic might occasionally explode and cause harm. Similarly, I would never take medications that hadn't been thoroughly examined for drug interactions, side effects, and safety.

Infermedica was established on core values of trust, transparency, quality, and courage. I firmly believe these values set the right mindset for an AI medical device manufacturer. Additionally, they form the foundation of a strong sense of responsibility—critical for deploying healthcare AI at scale.

The future of our product development

As we move towards our vision of "Infermedica 2.0," our goal is to leverage our strengths together with the capabilities of Generative AI and further expand our AI into a platform model. We aim to combine our thorough clinical validation and proven probabilistic reasoning with other models—especially fine-tuned frontier LLM models—under the concept of Neuro-Symbolic AI.

Our overarching ambition is to build autonomous or semi-autonomous agents capable of assisting or performing clinical tasks in a medically certified manner, starting with patient triage and patient navigation. This initial use case already shows strong demand from over 100 clients and partners we work with today.

The diagram above outlines key elements of our AI platform. In the Neuro-Symbolic approach, the "Neuro" component includes frontier LLMs, fine-tuned models, and ML-trained proprietary models built on top of 22 million patient interactions with Infermedica.

The "Symbolic" component consists of our human-curated knowledge graph and related ontology. To date, doctors have invested over 140,000 hours assembling, curating, and verifying our knowledge graphs using the highest quality, evidence-based literature. This allowed us to develop Bayesian reasoning algorithms integrated with LLMs, controlling clinical logic, eliminating hallucinations, and ensuring deterministic results. This approach provides complete explainability and transparency—crucial for safety, reliability, and the certification of a medical-grade product, which we aim to achieve later this year.

The agent orchestrator ensures models are used appropriately and in the right context, promoting principled AI reasoning that delivers explainable, factual, and probabilistic results, alongside providing the AI guardrails necessary for patient safety.

Moving forward, our goal is to expand the platform with new agent workflows and capabilities to support healthcare organizations in streamlining care delivery, optimizing resources, and improving patient outcomes, making Infermedica a trusted companion for common primary care interactions.

With this plan, we effectively combine the best aspects of symbolic and neural AI approaches. We believe this combination will create a human-like conversational and voice experience that can ultimately be certified as a medical device:

• Infermedica’s Strengths: Clinical validity, medical content curated through over 140,000 hours of physician work, medical device certification, deterministic and fully transparent reasoning that prevents hallucinations, and continuous human oversight.

• LLMs’ Strengths: Conversational interfaces that enhance patient language understanding, emulate empathy (compassion, recall), and integrate general clinical knowledge.

Beyond these aspects, the symbolic approach using knowledge graphs has another advantage: the ability to quickly and cost-effectively adapt new knowledge. Our knowledge graphs are maintained daily, with updates released every five weeks, and unlike Generative AI models, can be quickly investigated and revised through human review without the costly need to retrain the entire model.

Gen AI-powered Infermedica: Coming soon

What I’ve described isn’t just a vision or ambition—it’s a reality our exceptional team has been working on for the past 1.5 years. In the coming weeks, we'll reach a key milestone by launching pilots for our first Neuro-Symbolic product, Conversational Triage. We've meticulously built and iterated this product in close collaboration with our internal medical teams and through a usability study involving over 10,000 real users on our Symptomate platform. We're incredibly excited about the potential this new modality offers to enhance patient trust, understanding, and adoption.

We are also actively developing voice capabilities, now one of our key strategic initiatives. With successful deployments in call centers, such as the government-funded Healthdirect Australia and Medis as one of the largest private insurers in Portugal, we've gained invaluable insights for creating conversational agents that support nurses in their critical work of clinically evaluating and triaging patients.

But that's not all—company-wide, we plan to integrate Gen-AI-powered features into all components of Infermedica’s products in ways fully aligned with clinical validation and supervision.

If you’d like to learn more about what we’re up to – please reach out.