Multimodal AI for Musculoskeletal Diagnosis: Combining Imaging, Genetics and Patient Data

The limits of image-only AI

Most artificial intelligence in orthopedics today does one thing: it looks at a picture. Deep-learning models can grade knee osteoarthritis on an X-ray, segment cartilage on an MRI or flag a fracture on a radiograph with accuracy comparable to a specialist. These are real achievements, and they are already entering radiology workflows.

But a clinician does not decide from a picture alone. Two patients with identical Kellgren–Lawrence grade 2 knees can have completely different futures: one will stay stable for a decade, the other will need a joint replacement in five years. The difference lies in what the image does not show — genetic predisposition, inflammatory profile, body composition, activity level, pain sensitivity, comorbidities, and the patient's own account of how the joint behaves in daily life. Musculoskeletal disease is multifactorial, and a diagnostic AI that ignores most of the factors will plateau at "describing the picture" rather than predicting the patient.

This is the gap Spero Science AI was created to close.

What "multimodal" means in practice

Spero Science AI is a platform for multimodal musculoskeletal diagnostics: models that integrate three families of data that are usually analyzed in isolation.

Imaging. X-ray, MRI and ultrasound provide the structural picture — joint-space narrowing, osteophytes, bone-marrow lesions, cartilage thickness and composition. Quantitative MRI techniques such as T2 mapping go beyond morphology and measure the water and collagen organization of cartilage, an early marker of degeneration long before the surface breaks down. Spero's founders use exactly these techniques in their preclinical cartilage-regeneration studies, where T2 mapping distinguishes repaired tissue from native cartilage and from untreated defects.

Genetics and molecular data. Osteoarthritis, tendinopathy and cartilage injury all have a heritable component, and inflammatory and metabolic markers modulate how a joint responds to load and to treatment. Adding genomic and biomarker layers lets a model separate patients who look alike on imaging but differ in biology.

Patient-reported and clinical data. Validated outcome instruments (pain, function, quality of life), activity profiles, anthropometrics, surgical history and comorbidities capture the lived reality of the disease — and they are the outcomes that matter to patients and to health systems.

A multimodal model learns from all three at once. Instead of "what grade is this knee?", it can address the questions that change decisions: Will this patient's cartilage lesion progress within two years? Is this patient likely to respond to an intra-articular biologic, or should surgery be considered now? Which rehabilitation profile fits this athlete's return-to-sport risk?

Precision orthopedics: from population averages to individual decisions

Orthopedics has historically treated patients by category — lesion size, age band, disease grade. Precision medicine replaces the category with the individual. The goal of Spero Science AI is to give surgeons and physiatrists a decision-support layer that estimates, for a specific patient, the probability of progression and the expected benefit of each therapeutic option, from conservative care to regenerative therapies to surgery.

That objective connects directly to Spero's therapeutic platforms. A chondroinductive injection or an engineered cell construct will work best in the right patient at the right stage of disease. Multimodal diagnostics are the tool for finding that patient — for selecting trial participants, stratifying responders and, eventually, guiding treatment in routine practice. Diagnostics and therapeutics are designed as one system.

Built by clinicians, on clinical data

Spero Science AI is led by orthopedic surgeon-scientists from the University of São Paulo Medical School (FMUSP) and Hospital das Clínicas HC-FMUSP, one of the largest hospital complexes in Latin America, with the sports-medicine practice of Hospital Sírio-Libanês. The founding team combines knee surgery and sports medicine, cartilage biology and tissue engineering with an active research program in the musculoskeletal sciences graduate program at USP.

This matters for two reasons. First, the models are trained to answer questions that clinicians actually face, using outcome definitions that clinicians accept. Second, the Brazilian population is one of the most genetically diverse in the world, and models developed on it are less likely to inherit the biases of datasets drawn from a single ancestry — an increasingly important consideration for AI that will be used internationally.

Responsible AI in healthcare

Spero Science develops its AI under the same principles that govern clinical research: ethics-committee approval, informed consent, de-identification and data protection under Brazil's LGPD, external validation before any clinical use, and transparent reporting of performance across subgroups. Models are positioned as decision support for qualified professionals, not as autonomous diagnosis. Regulatory classification as software as a medical device (SaMD) is planned in line with ANVISA requirements and international guidance.

Development status

Spero Science AI is in development: assembling multimodal datasets, building and validating predictive models for osteoarthritis and cartilage-injury progression, and defining the clinical validation studies that will support regulatory submission. The platform also serves as the analytics backbone for Spero's therapeutic programs, from patient selection in preclinical-to-clinical translation to outcome measurement.

Working with us

Hospitals and research groups with musculoskeletal imaging and outcome data, genomics companies, medical-device and pharmaceutical partners, and investors focused on health AI can collaborate with Spero Science on data partnerships, validation studies, co-development and licensing.

Learn more about Spero Science AI, see how it connects with our Chondroinductive Molecule and Smart Cells platforms, or contact the team.

References

• Fernandes TL, Santanna JPC, de Faria RR, Pastore ER, Bueno DF, Hernandez AJ. Tissue Engineering Construct for Articular Cartilage Restoration with Stromal Cells from Synovium vs. Dental Pulp — A Pre-Clinical Study. Pharmaceutics. 2024;16(12):1558. https://doi.org/10.3390/pharmaceutics16121558 (quantitative MRI / T2 mapping of repaired cartilage)

• Hinckel BB, Thomas D, Vellios EE, et al. (incl. Fernandes TL). Algorithm for Treatment of Focal Cartilage Defects of the Knee: Classic and New Procedures. Cartilage. 2021;13(1_suppl):473S-495S. https://doi.org/10.1177/1947603521993219 (decision criteria that precision diagnostics aim to personalize)