Introducing Sylvester’s updated AI model

Our updated computer vision model is exceptionally good at identifying cats who are not in discomfort. That is the property that makes our Cat Comfort Check most clinically useful, and the story behind a release that improved every metric in our evaluation suite.

Negative predictive value

The updated model achieves a negative predictive value, or NPV, of 90.3%. NPV measures how reliable a negative result is. When a Cat Comfort Check returns a comfortable result, that result is correct roughly nine times out of ten.

For veterinarians, this is the property that matters most. A screening tool earns its place by meaningfully narrowing the field of cats who require further investigation, and a high NPV is exactly what does that. It's the same logic that makes SNAP and IDEXX panels useful in your daily practice.

The wider performance picture

Specificity, the model's ability to correctly identify cats who are not in discomfort, improved 8.6%. Positive likelihood ratio, which captures how strongly a positive result indicates genuine discomfort, climbed 51.3%.

Precision, which measures how reliably a positive result corresponds to genuine discomfort, rose 29.6%. Sensitivity, the rate at which the model catches cats genuinely in discomfort, improved 4.6%.

Overall accuracy improved 7.7%, and F1 score, the balance of precision and sensitivity, climbed 17.3%.

Reading a result

Sylvester is a caregiver screening tool, not a diagnostic device. Veterinary diagnosis remains the responsibility of qualified clinicians.

A comfortable result is clinically meaningful. The high NPV means it substantially reduces the probability that a cat is in discomfort at the time of the check. It doesn't rule pain out entirely, and standard veterinary follow-up remains appropriate where other clinical signs are present.

An uncomfortable result is a prompt for evaluation, not a diagnosis. Precision improved 29.6% in this update and continues to climb. Because the model is tuned for screening, we deliberately prioritise catching every cat in discomfort over filtering out false positives, since a missed pain case is a more serious error than one resolved on examination. Every uncomfortable result is designed to trigger a clinical conversation, which is where diagnosis happens.

Where the updated model sits

We benchmark our models against an internal four-tier framework that reflects clinical readiness. Tier 0 is research and exploratory. Tier 1 is a screening prototype meeting indicative thresholds for sensitivity, specificity, and predictive value. Tier 2 is a preclinical candidate suitable for formal prospective evaluation. Tier 3 is clinical decision-support, suitable to influence clinical decisions in prospective studies.

The updated model qualifies as a strong Tier 1 prototype, with specificity already at our Tier 2 benchmark. Sensitivity, precision, and balanced accuracy remain below Tier 2 thresholds and are the focus of ongoing research.

Research and training data

The updated model reflects an expanded training dataset and additional data licensing agreements with veterinary partners, shelters and research collaborators. The annotation framework draws on Sylvester's feline discomfort assessment protocol, which has now been validated and builds on established feline pain literature including Brondani et al. 2013, Calvo et al. 2014, and Evangelista et al. 2019.

Our research team, led by Lead Animal Behaviourist Frances Valentine, PhD, is studying cats across a range of health conditions including oral, orthopaedic, respiratory, and wound-related discomfort, alongside how physical variables like coat length, coat colour, face shape, and body condition score affect signal interpretation. Posture as a complementary indicator to facial expression is also an active research area.

Closing the gap between home and clinic

70% of cats do not receive regular veterinary care, and many present at a clinic only once a condition has advanced. Sylvester was built to close that feline discomfort awareness gap. Caregivers use it at home to run a Cat Comfort Check, where the cat is most comfortable, on their own schedule. When a check returns an uncomfortable result, a caregiver who might otherwise have waited weeks, or never come in at all, has a clear, validated reason to call a vet.

The pathway is already showing up in the numbers. According to early user data 30% of Cat Comfort Checks lead to a vet visit, and 80% of those visits result in diagnostics or treatment delivered by the attending clinician. That's care reaching cats who, in many cases, weren't going to come in this year.

Veterinary Clinic Partnership

Sylvester gives caregivers objective insight into their cat's comfort at home. Instead of relying on a sense that something seems off, they have an at-home observation they can act on. The Symptom Checker walks caregivers through next steps when a Cat Comfort Check returns an uncomfortable result, and often that step is to call your practice. Between visits, caregivers using Sylvester are checking in on their cat's comfort regularly, building a pattern of at-home observations over time. A contact your clinic button will let caregivers reach your practice directly from a Cat Comfort Check result.

Practice Portal access gives your clinic visibility into how connected patients are doing between exam days, supporting earlier and better-informed conversations when they walk through your door. Every Cat Comfort Check keeps your practice top of mind. Caregivers connected to your clinic see your name throughout their everyday care, reinforcing the connection between their cat's wellbeing and your team. The compounding effect is client retention. Caregivers who feel supported between appointments stay loyal, and Sylvester gives your practice a low-effort way to remain present in your clients' lives.

For clinics looking to see more cats or better engage their existing client base, we provide a comprehensive partnership package including caregiver-facing materials like flyers, social templates, and educational blogs.

Limitations and roadmap

We're transparent about what the updated model doesn't yet do. Some cats in discomfort will still be missed at the default screening threshold, brachycephalic breeds such as Persians and Himalayans see lower expected accuracy, and the model is not validated for cats 2 months and under. The next update targets Tier 2 status with a primary focus on sensitivity, precision, and balanced accuracy, supported by multi-rater validation, an expanded training dataset, and continued research across diverse cat populations.

Contact Us

We are actively building partnerships with veterinary clinics, researchers, and data contributors.

Get in touch: hello@sylvester.ai.