AI in Orthopedics

A family of algorithms that learn patterns from data rather than being explicitly programmed with rules. In orthopedics, machine learning models are commonly trained on labeled imaging or outcomes data.

A subset of machine learning that uses layered neural networks, capable of learning complex features directly from raw data such as radiographs or MRI scans. Most recent AI imaging tools in orthopedics rely on deep learning.

A deep learning architecture designed for image data. CNNs scan an image with small filters to detect local features such as edges and textures, and remain a standard architecture for fracture detection, segmentation, and OA grading.

A neural network architecture built around an attention mechanism that weighs relationships across an input sequence. Transformers power large language models and are increasingly used for medical imaging and clinical text.

A large model pretrained on a broad dataset that can be adapted to many downstream tasks. Foundation models are often the backbone of newer clinical AI tools and include families such as GPT, Claude, and medical-specific variants.

A transformer-based foundation model trained on text. LLMs produce fluent natural-language output and can summarize literature, draft notes, and answer clinical questions, but they can also fabricate information. See Hallucination.

The natural-language input given to a large language model. Prompt wording substantially affects output quality, and small changes to a prompt can change the model’s answer.

The production by a large language model of confident-sounding content that is false or unsupported by any source. Hallucinations are a well-documented failure mode and a central safety concern for clinical LLM use.

A technique in which a language model is connected to a curated document store at query time and instructed to answer from retrieved passages. RAG is used to reduce hallucinations in clinical and research tools.

The proportion of true positives correctly identified by a test or model. A fracture detection model with 95 percent sensitivity misses 5 percent of fractures present in the data.

The proportion of true negatives correctly identified by a test or model. High sensitivity with low specificity produces many false alarms. Both numbers should always be reported together.

Among cases the model flags as positive, the proportion that are truly positive. PPV depends on disease prevalence and falls sharply in low-prevalence populations.

Among cases the model flags as negative, the proportion that are truly negative. NPV also depends on prevalence and is often high when disease is uncommon.

A receiver operating characteristic curve plots sensitivity against 1 minus specificity across all classification thresholds. The area under the curve (AUC) summarizes overall discrimination on a 0 to 1 scale, with 0.5 indicating chance and 1.0 indicating perfect discrimination.

The three data partitions used to develop and evaluate a model. The model learns from the training set, is tuned on the validation set, and is evaluated on the held-out test set. A model that has seen the test data during training will report inflated performance.

When a model learns patterns specific to its training data that do not generalize to new cases. Overfit models perform well on their own test set but fail on external cases.

Evaluation of a model on data from an institution, scanner, or population not used during training. External validation is the standard for judging whether a model will generalize to clinical use.

The reference label against which model predictions are compared, for example an orthopedic surgeon’s read of a radiograph or a confirmed intraoperative diagnosis. Model performance is only as reliable as the ground truth it is measured against.

Picture Archiving and Communication System. The hospital infrastructure that stores, retrieves, and distributes medical imaging. Clinical AI imaging tools are typically integrated at the PACS level.

Digital Imaging and Communications in Medicine. The standard file format and communication protocol for medical imaging. AI imaging models typically consume DICOM inputs.