A patient presents to the emergency department with a spreading soft tissue infection and fever. The orthopaedic team is asked to review. Inflammatory markers are elevated, the limb is swollen and tender, and the clinical picture is consistent with serious deep infection — but the critical question is not yet answered: is this necrotizing fasciitis, requiring emergency surgical debridement, or osteomyelitis, which can initially be managed with intravenous antibiotics and a more deliberate surgical decision?
This distinction matters acutely. Necrotizing fasciitis is a rapidly progressive, potentially fatal infection where surgical delay measurably worsens outcomes. Osteomyelitis is serious but survivable with a different — and less urgent — management pathway. The clinical problem is that in the early stages, these two conditions can look nearly identical. It is not a deficiency in clinical skill. It is an inherent feature of how these infections present, and it is exactly the kind of high-stakes diagnostic problem that machine learning is well positioned to address.
The paper
Yasin et al. (2026, NPJ Digital Medicine; PMID 42056542) developed and externally validated an explainable machine learning model designed to distinguish necrotizing fasciitis from osteomyelitis using routine blood biomarkers. The study draws on a retrospective, multicentre cohort of 3,415 patients — 579 with NF and 2,836 with OM — making this one of the larger datasets yet assembled for this diagnostic problem. Data from a primary centre were used for model development, with an independent second centre providing the external test cohort.
Systematic evaluation identified an optimal 10-biomarker LightGBM model. Discriminative performance on external validation was strong: AUC 0.926. The model’s explainability analyses confirmed predictions are driven by clinically relevant markers of severe inflammation and metabolic dysfunction — not artefacts of the training data. The model has been deployed as a publicly accessible web tool for real-time risk stratification at the bedside.
Why this matters on-call
Existing clinical decision rules for soft tissue infections were not designed to reliably differentiate NF from OM across the clinical spectrum, and in patients with NF, the classic signs of systemic sepsis can lag behind local tissue destruction. A patient can look less unwell than the limb is. CRP may be rising but not dramatically; white cell count may be only modestly elevated; fever may be low-grade. The window in which confident clinical differentiation is possible on examination alone can be narrow.
A model that integrates biomarker profiles specifically associated with the pathophysiology of necrotizing fasciitis — rather than just the generic systemic inflammatory response — has a plausible mechanism for outperforming simple clinical rules. AUC 0.926 at external validation in a 3,415-patient cohort is genuinely strong performance. The dataset size is particularly noteworthy: NF studies are often constrained by the relative rarity of the condition, which makes this multicentre approach a meaningful step forward in terms of generalisability.
The web-deployed tool
Deploying the model as a freely accessible web tool addresses the access problem without requiring EPR integration. Any clinician with a smartphone can enter the available blood values at the bedside and receive an output in real time.
The practical limitation is that web deployment without pathway integration depends on the individual clinician knowing the tool exists, choosing to use it, and correctly interpreting the output. This is the difference between a tool embedded in care and a tool that requires initiative to invoke. Both have value — but the latter is more vulnerable to the conditions of a busy overnight take. A high-probability output for NF in a patient who doesn’t yet look critically unwell may be exactly what supports an escalation call that otherwise feels premature.
What to do with this
For the trainee covering orthopaedic on-call: this model will not change your immediate management of a sick patient with suspected NF. If NF is in the differential, you are escalating, planning for theatre, and getting senior input regardless. What a high-probability ML output adds is a quantitative flag in the conversation — something to cite when making the case for urgent surgical review at 02:00 when the patient doesn’t yet look as unwell as the diagnosis demands.
For the consultant: this is a good template for what well-designed clinical AI looks like. A specific, high-stakes diagnostic problem. A large, multicentre dataset. External validation. Explainability built in. Freely available. Compare that template to the generic “AI-enhanced” labels that populate procurement catalogues, and the gap is clear.
The condition that cannot wait has always needed the best available decision support. A freely accessible, externally validated ML tool for NF differentiation — now deployed online — is worth knowing about before you need it.
References
- Yasin P, et al. Explainable machine learning differentiates necrotizing fasciitis and osteomyelitis via routine blood biomarkers. NPJ Digit Med. 2026. PMID 42056542. https://doi.org/10.1038/s41746-026-02686-3