AI & ML interests
None defined yet.
Recent Activity
Medrotic Research
Advancing Trustworthy AI for Healthcare
Medical imaging ยท Vision-language models ยท Reproducible machine learning
About Medrotic Research
Medrotic Research is the research-focused initiative connected with Medrotic โ AI Healthcare. We explore machine learning, medical computer vision, and multimodal AI, with an emphasis on transparent evaluation, reproducibility, and responsible use.
Our work is experimental. Research models and results presented here should not be interpreted as clinically approved diagnostic technology.
Medrotic โ AI Healthcare Platform
Visit the live Medrotic project
Medrotic is our collaborative AI-powered healthcare software project. This Hugging Face organization provides a home for related research, documented experiments, and possible future model or dataset releases.
The live healthcare platform and the research experiments are distinct. A model described here is not necessarily integrated into the deployed platform.
Featured Research: Medrotic-TB
Research area: Classification of chest radiographs with appearances suggestive of pulmonary tuberculosis
Status: Experimental research in progress
Primary dataset: NLM Shenzhen chest X-ray dataset
Medrotic-TB investigates resource-efficient adaptation and evaluation of medical vision-language models, alongside specialist medical-image representations and conventional classification baselines.
Research tracks
| Research track | Approach |
|---|---|
| Medical vision-language models | MedGemma and parameter-efficient adaptation using QLoRA |
| Medical image representations | MedSigLIP embeddings and linear-probe evaluation |
| Comparison methods | DenseNet-121 frozen image representations |
| Robustness and uncertainty | Controlled perturbations and selective-prediction analysis |
| Research quality | Dataset provenance, frozen partitions, reproducibility checks, audit trails, and error analysis |
Experimental milestones
| Experiment | Status | Scope |
|---|---|---|
| E1-2 โ MedGemma adaptation | Completed | Rank-4 QLoRA adaptation using the frozen 430-case training partition |
| E2-F1 โ MedSigLIP baseline | Completed | Frozen image embeddings and logistic-regression evaluation |
| E2-F2 โ DenseNet-121 baseline | Completed | Frozen ImageNet-pretrained image representations |
| E3-F2 โ Reproducibility replay | Completed | Development-set prediction replay |
| E4-F1 โ Robustness analysis | Completed | Controlled perturbation experiments on the development partition |
| E4-F2 โ Selective prediction | Completed | Development-set uncertainty and selective-prediction analysis |
| E5-1 โ Calibration | Completed; archive self-audit passed | Pre-specified calibration procedures on 66 Shenzhen calibration cases |
E5-1 calibration update โ 18 November 2026
The E5-1 experiment was completed under MEDROTIC-TB Protocol Amendment v1.1.
It evaluated pre-specified fixed-reference and sensitivity-oriented thresholds, as well as selective-prediction cutoffs, for MedGemma E1-2 and MedSigLIP E2-F1.
The calibration partition contained 66 cases:
- 34 with radiographic appearances suggestive of TB.
- 32 normal cases.
The sensitivity-oriented calibration procedure produced the following calibration-set estimates:
| Model | Sensitivity | Specificity | Accuracy |
|---|---|---|---|
| MedGemma E1-2 | 91.18% | 96.88% | 93.94% |
| MedSigLIP E2-F1 | 91.18% | 90.63% | 90.91% |
These figures are calibration-set results, not independent test performance. They must not be interpreted as evidence of external generalization, clinical diagnostic accuracy, or guaranteed sensitivity on unseen data.
The E5-1 experiment did not retrain or modify the MedGemma adapter. Its calibration thresholds and analysis artifacts are maintained separately from the model weights.
Evaluation boundaries
- The study uses a frozen Shenzhen v6.1 partitioning scheme.
- The 67-case internal test partition remains sealed; no internal-test results are reported.
- The Montgomery cohort has been excluded from the current study.
- Independent clinical-reader validation has not been performed.
- Calibration estimates are based on a small, single-dataset partition.
- External validation, clinical utility, and deployment readiness have not been established.
The classification target is radiographic appearances suggestive of pulmonary tuberculosis, not microbiologically confirmed active TB.
Development and calibration results do not establish clinical performance or generalizability.
Research Interests
- Medical imaging and computer vision
- Multimodal AI and vision-language models
- Efficient fine-tuning and model adaptation
- Robustness, explainability, and trustworthy AI
- Uncertainty estimation and selective prediction
- Reproducible experimentation and benchmark evaluation
Team Medrotic
Medrotic was developed collaboratively by three team members:
| Team member | Involvement |
|---|---|
| Muhammad Ibtisam Afzal | AI/ML research, frontend and backend development, and AI integration; GitHub |
| Muhammad Adnan Tariq | AI/ML lead for the Medrotic platform: MedGemma, RAG chatbot (LangChain, ChromaDB), prescription OCR, and the FastAPI AI service; Google Scholar |
| Muhammad Khaqan Nasir | Medrotic project team member |
Academic FYP supervisor: Dr. Muhammad Farhan, Department of Computer Science, COMSATS University Islamabad, Sahiwal Campus.
Our Research Principles
- Evidence first: Report results, experimental limitations, and uncertainty accurately.
- Reproducibility: Track data provenance, model versions, configurations, and evaluation procedures.
- Responsible development: Do not equate research prototypes with clinically validated systems.
- Transparent evaluation: Separate training, development, calibration, and sealed test data.
- Open collaboration: Share code, methods, and artifacts where permissions and quality standards allow.
Explore & Connect
- Live project: medrotic.vercel.app
- Research updates: Follow the Medrotic Research organization here on Hugging Face.
- Research adapter: MedGemma E1-2 adapter
Medical disclaimer: All experimental AI models discussed here are for research purposes only. They are not intended for clinical diagnosis, treatment recommendations, patient management, or medical decision-making.