Lymphoedema Diagnosis and Treatment

Artificial Intelligence (AI), Lymphedema

lymphedema, chat gpt, artificial intelligence

A domain-specific, custom-trained large language model for the differential diagnosis and treatment planning of lymphedema, lipedema, and venous insufficiency.

The differential diagnosis of lower limb swelling remains problematic in clinical practice, as lymphedema, lipedema, and peripheral venous disease often present with similar features. Therefore, we developed LymphedemaGPT, a GPT-5-based clinical assistant designed to help practitioners navigate these diagnostic complexities. LymphedemaGPT was designed to analyze structured patient data to extract clinical summaries, present possible diagnoses with percentage probabilities, create differential diagnosis tables, suggest additional diagnostic tests, and generate evidence-based treatment plans. LymphedemaGPT's responses are based on seven scientific publications uploaded to the system, in addition to the Sleigh BC & Manna B (2023) and Rockson approaches. Owing to this resource integration, the model can provide more reliable and consistent recommendations aligned with evidence-based medicine principles based on current guidelines and scientific publications. Extensive prompt engineering techniques were applied to optimize the diagnostic and therapeutic accuracy of LymphedemaGPT. The model is programmed to prioritize the questioning phase until a diagnosis is confirmed. In the initial responses, only structured anamnesis questions were asked, and after sufficient information was collected, systematic analysis and treatment planning were initiated. The response flow was designed as follows: (1) history collection, (2) preliminary assessment, (3) additional questioning (if necessary), and (4) systematic analysis and treatment planning when sufficient data were obtained. The following patient data was presented to LymphedemaGPT in a structured format: Demographic data: Age, gender, height, weight Medical history: Additional illnesses, medications used, habits (smoking, alcohol) Complaint characteristics: Time of onset, affected area, symptoms (pain, heaviness, numbness, tingling, stiffness, limited movement, weakness, etc.) Physical examination findings: Stemmer sign, swelling change with elevation, skin findings Medical history: History of infection, history of surgery, history of malignancy (radiotherapy, chemotherapy, lymph node dissection, type of cancer) Imaging: Doppler ultrasonography and lymphoscintigraphy results, if available LymphedemaGPT was asked to respond in the following 12-part standard format: (1) Clinical Summary, (2) Possible Diagnoses (% probability), (3) Differential Diagnosis, (4) Recommended Diagnostic Tests, (5) Treatment Plan, (6) Patient Education and Follow-up, (7) Red Flags, (8) references, (9) Level of Evidence and Confidence Score, (10) Ethical Note, (11) Data Summary (JSON/CSV), and (12) Analysis Timestamp. The performance of LymphedemaGPT was evaluated by experienced physicians based on the following eight criteria: 1. Accuracy and adequacy of clinical summary 2. Accuracy of primary diagnosis 3. Accuracy of differential diagnosis table 4. Appropriateness of recommended diagnostic tests 5. Concordance of treatment plan with current guidelines 6. Appropriateness of compression class/exercise-diet recommendations 7. Adequacy of red flags 8. Overall clinical utility Each criterion was scored using a 5-point Likert scale: 5 = excellent/completely suitable, 4 = good/significantly suitable, 3 = moderate/partially suitable, 2 = poor/inadequate, and 1 = very poor/not suitable. The maximum score for each case was 40 (8 criteria × 5 points), and the minimum score was 8. Two experienced physicians independently performed the evaluation. The evaluators were physical medicine and rehabilitation specialists experienced in the management of lymphedema and lipedema, and they independently performed the scoring.

No related papers found yet.

Turkey (Türkiye)