ghndrx/kubeflow-pipelines
1
0
Fork 0
mirror of https://github.com/ghndrx/kubeflow-pipelines.git synced 2026-02-10 06:45:13 +00:00
Code Issues Packages Projects Releases Wiki Activity

docs: clean up README and USE_CASES formatting

Browse Source
This commit is contained in:
Greg Hendrickson
2026-02-03 17:07:07 +00:00
parent 0bf3837e78
commit 210d9c8999
2 changed files with 107 additions and 174 deletions
Download Patch File Download Diff File Expand all files Collapse all files

178
USE_CASES.md
View File

@@ -1,158 +1,92 @@
# Healthcare ML Use Cases & Datasets
Curated list of similar healthcare/biomedical use cases with publicly available datasets for training on RunPod.
Curated list of healthcare/biomedical use cases with publicly available datasets.
---
## 🔥 Priority 1: Ready to Train
## Implemented
### 1. Adverse Drug Event Classification
**Dataset:** `Lots-of-LoRAs/task1495_adverse_drug_event_classification`
- **Task:** Classify text for presence of adverse drug events
- **Size:** ~10K samples
- **Labels:** Binary (adverse event / no adverse event)
- **Use Case:** Pharmacovigilance, FDA reporting automation
- **Model:** Bio_ClinicalBERT
### 1. Drug-Drug Interaction (DDI) Classification
- **Dataset:** DrugBank (bundled)
- **Task:** Classify interaction severity
- **Size:** 176K samples
- **Labels:** Minor, Moderate, Major, Contraindicated
- **Status:** Production ready
```python
from datasets import load_dataset
ds = load_dataset("Lots-of-LoRAs/task1495_adverse_drug_event_classification")
```
### 2. PubMed Multi-Label Classification (MeSH)
**Dataset:** `owaiskha9654/PubMed_MultiLabel_Text_Classification_Dataset_MeSH`
- **Task:** Assign MeSH medical subject headings to research articles
- **Size:** ~50K articles
- **Labels:** Multi-label (medical topics)
- **Use Case:** Literature categorization, research discovery
- **Model:** PubMedBERT
```python
from datasets import load_dataset
ds = load_dataset("owaiskha9654/PubMed_MultiLabel_Text_Classification_Dataset_MeSH")
```
### 2. Adverse Drug Event Detection
- **Dataset:** `ade-benchmark-corpus/ade_corpus_v2`
- **Task:** Binary classification for ADE presence
- **Size:** 30K samples
- **Labels:** ADE / No ADE
- **Status:** Production ready
### 3. Symptom-to-Disease Prediction
**Dataset:** `shanover/disease_symptoms_prec_full`
- **Task:** Predict disease from symptom descriptions
- **Size:** Variable
- **Labels:** Disease categories
- **Use Case:** Triage, symptom checker apps
- **Model:** Bio_ClinicalBERT
```python
from datasets import load_dataset
ds = load_dataset("shanover/disease_symptoms_prec_full")
```
- **Dataset:** `shanover/disease_symptoms_prec_full`
- **Task:** Predict disease from symptoms
- **Size:** ~5K samples
- **Labels:** 41 disease categories
- **Status:** Production ready
### 4. Medical Triage Classification
**Dataset:** `shubham212/Medical_Triage_Classification`
- **Task:** Classify urgency level of medical cases
- **Size:** ~500 downloads (popular)
- **Labels:** Triage levels (Emergency, Urgent, Standard)
- **Use Case:** ER automation, telemedicine routing
- **Model:** Bio_ClinicalBERT
- **Dataset:** `shubham212/Medical_Triage_Classification`
- **Task:** Classify urgency level
- **Labels:** Emergency, Urgent, Standard, Non-urgent
- **Status:** Production ready (needs more training data)
---
## 📚 Priority 2: QA & Reasoning
## Future Candidates
### 5. MedMCQA - Medical Exam Questions
**Dataset:** `openlifescienceai/medmcqa` (24K downloads!)
### PubMed Multi-Label Classification (MeSH)
- **Dataset:** `owaiskha9654/PubMed_MultiLabel_Text_Classification_Dataset_MeSH`
- **Task:** Assign MeSH subject headings to articles
- **Size:** 50K articles
- **Use Case:** Literature categorization
### MedMCQA - Medical Exam QA
- **Dataset:** `openlifescienceai/medmcqa`
- **Task:** Answer medical entrance exam questions
- **Size:** 194K MCQs covering 2.4K healthcare topics
- **Labels:** Multiple choice (A/B/C/D)
- **Size:** 194K MCQs
- **Use Case:** Medical education, knowledge testing
- **Model:** Llama-3 or Gemma (LLM fine-tuning)
```python
from datasets import load_dataset
ds = load_dataset("openlifescienceai/medmcqa")
```
### 6. PubMedQA - Research Question Answering
**Dataset:** `qiaojin/PubMedQA` (18K downloads!)
- **Task:** Answer yes/no/maybe questions from abstracts
### PubMedQA - Research Question Answering
- **Dataset:** `qiaojin/PubMedQA`
- **Task:** Yes/No/Maybe from abstracts
- **Size:** 274K samples
- **Labels:** yes / no / maybe
- **Use Case:** Evidence-based medicine, literature review
- **Model:** PubMedBERT or Bio_ClinicalBERT
- **Use Case:** Evidence-based medicine
```python
from datasets import load_dataset
ds = load_dataset("qiaojin/PubMedQA")
```
### Medical Abbreviation Disambiguation
- **Dataset:** `McGill-NLP/medal`
- **Task:** Disambiguate abbreviations in context
- **Size:** 4GB curated
- **Use Case:** Clinical note processing
---
## 🧬 Priority 3: Specialized NLP
### 7. Medical Abbreviation Disambiguation (MeDAL)
**Dataset:** `McGill-NLP/medal`
- **Task:** Disambiguate medical abbreviations in context
- **Size:** 14GB → curated to 4GB
- **Labels:** Abbreviation meanings
- **Use Case:** Clinical note processing, EHR parsing
- **Model:** Bio_ClinicalBERT
### 8. BioInstruct - Instruction Following
**Dataset:** `bio-nlp-umass/bioinstruct`
### BioInstruct
- **Dataset:** `bio-nlp-umass/bioinstruct`
- **Task:** Instruction-tuned biomedical tasks
- **Size:** 25K instructions
- **Labels:** Various biomedical tasks
- **Use Case:** General biomedical assistant
- **Model:** Llama-3 or Mistral (LoRA fine-tuning)
---
## 🛠️ Implementation Roadmap
## Dataset Comparison
### Week 1: Adverse Drug Events
1. Download ADE dataset
2. Add to handler.py as new training mode
3. Train classifier → S3
4. Build inference endpoint
### Week 2: PubMed Classification
1. Download PubMed MeSH dataset
2. Multi-label classification head
3. Train → S3
4. Literature search API
### Week 3: Medical QA
1. Download MedMCQA
2. LLM fine-tuning with LoRA
3. Deploy QA endpoint
### Week 4: Symptom Checker
1. Symptom-disease dataset
2. Train classifier
3. Build symptom input → disease prediction API
| Dataset | Size | Task | Complexity |
|---------|------|------|------------|
| DDI (DrugBank) | 176K | 4-class | Medium |
| ADE Corpus | 30K | Binary | Low |
| PubMed MeSH | 50K | Multi-label | High |
| MedMCQA | 194K | MCQ | High |
| PubMedQA | 274K | 3-class | Medium |
| Symptom-Disease | 5K | 41-class | Medium |
| Triage | 5K | 4-class | Low |
---
## 📊 Dataset Comparison
## Additional Resources
| Dataset | Size | Task | Difficulty | Business Value |
|---------|------|------|------------|----------------|
| DDI (current) | 176K | Classification | Medium | ⭐⭐⭐⭐⭐ |
| Adverse Events | 10K | Binary | Easy | ⭐⭐⭐⭐⭐ |
| PubMed MeSH | 50K | Multi-label | Medium | ⭐⭐⭐⭐ |
| MedMCQA | 194K | MCQ | Hard | ⭐⭐⭐⭐ |
| PubMedQA | 274K | Yes/No/Maybe | Medium | ⭐⭐⭐⭐ |
| Symptom→Disease | Varies | Classification | Easy | ⭐⭐⭐⭐⭐ |
| Triage | ~5K | Classification | Easy | ⭐⭐⭐⭐⭐ |
---
## 🔗 Additional Resources
- **MIMIC-III/IV:** ICU clinical data (requires PhysioNet access)
- **MIMIC-III/IV:** ICU clinical data (PhysioNet access required)
- **n2c2 Challenges:** Clinical NLP shared tasks
- **i2b2:** De-identified clinical records
- **ChemProt:** Chemical-protein interactions
- **BC5CDR:** Chemical-disease relations
---
*Generated: 2026-02-03*