How Machine Learning and AI can support the fight against COVID

Presented at NDI Conf and DataArt IT Non-stop

Transcript

1. How Machine Learning and AI can support the fight against COVID-19 Dmitry Soshnikov, PhD Senior Cloud Advocate, Microsoft Associate Professor, HSE/MIPT/MAI @shwars
2. Problem Around 30,000 scientific papers related to COVID appear monthly
3. CORD Papers Dataset Data Source https://allenai.org/data/cord-19 https://www.kaggle.com/allen-institute-for-ai/CORD-19- research-challenge CORD-19 Dataset Contains over 800,000 scholarly articles about COVID-19 and the coronavirus family of viruses for use by the global research community 400,000+ articles with full text
4. Natural Language Processing Common tasks for NLP: • Intent Classification • Named Entity Recognition (NER) • Keyword Extraction • Text Summarization • Question Answering • Open Domain Question Answering Language Models: • Recurrent Neural Network (LSTM, GRU) • Transformers • GPT-2 • BERT • Microsoft Turing-NLG • GPT-3 Microsoft Learn Module: Introduction to NLP with PyTorch aka.ms/pytorch_nlp docs.microsoft.com/en-us/learn/paths/pytorch-fundamentals/
5. How BERT Works (Simplified) Masked Language Model + Next Sentence Prediction During holidays, I like to ______ with my dog. It is so cute. 0.85 Play 0.05 Sleep 0.09 Fight 0.80 YES 0.20 NO BERT contains 345 million parameters => very difficult to train from scratch! In most of the cases it makes sense to use pre-trained language model.
6. Main Idea Use NLP tools to extract semi-structured data from papers, to enable semantically rich queries over the paper corpus. Extracted JSON Cosmos DB Database Power BI Dashboard SQL Queries Azure Semantic Search NER Relations Text Analytics for Health CORD Corpus
7. Part 1: Extracting Entities and Relations Base Language Model Dataset Kaggle Medical NER: • ~40 papers • ~300 entities Generic BC5CDR Dataset • 1500 papers • 5000 entities • Disease / Chemical Generic BERT Model Pre-training BERT on Medical texts PubMedBERT pre-trained model by Microsoft Research Huggingface Transformer Library: https://huggingface.co/
8. 6794356|t|Tricuspid valve regurgitation and lithium carbonate toxicity in a newborn infant. 6794356|a|A newborn with massive tricuspid regurgitation, atrial flutter, congestive heart failure, and a high serum lithium level is described. This is the first patient to initially manifest tricuspid regurgitation and atrial flutter, and the 11th described patient with cardiac disease among infants exposed to lithium compounds in the first trimester of pregnancy. Sixty-three percent of these infants had tricuspid valve involvement. Lithium carbonate may be a factor in the increasing incidence of congenital heart disease when taken during early pregnancy. It also causes neurologic depression, cyanosis, and cardiac arrhythmia when consumed prior to delivery. 6794356 0 29 Tricuspid valve regurgitation Disease D014262 6794356 34 51 lithium carbonate Chemical D016651 6794356 52 60 toxicity Disease D064420 6794356 105 128 tricuspid regurgitation Disease D014262 6794356 130 144 atrial flutter Disease D001282 6794356 146 170 congestive heart failure Disease D006333 6794356 189 196 lithium Chemical D008094 6794356 265 288 tricuspid regurgitation Disease D014262 6794356 293 307 atrial flutter Disease D001282 6794356 345 360 cardiac disease Disease D006331 6794356 386 393 lithium Chemical D008094 6794356 511 528 Lithium carbonate Chemical D016651 6794356 576 600 congenital heart disease Disease D006331
9. NER as Token Classification Tricuspid valve regurgitation and lithium carbonate toxicity in a newborn infant. Tricuspid B-DIS valve I-DIS regurgitation I-DIS and O lithium B-CHEM carbonate I-CHEM toxicity B-DIS in O a O newborn O infant O . O
10. PubMedBert, Microsoft Research from transformers import AutoTokenizer, BertForTokenClassification, Trainer mname = “microsoft/BiomedNLP-PubMedBERT-base- uncased-abstract” tokenizer = AutoTokenizer.from_pretrained(mname) model = BertForTokenClassification .from_pretrained(mname, num_labels=len(unique_tags)) trainer = Trainer( model=model, args=training_args, train_dataset=train_dataset, eval_dataset=val_dataset) trainer.train()
11. Notebooks Automated ML UX Designer Reproducibility Automation Deployment Re-training CPU, GPU, FPGAs IoT Edge Azure Machine Learning Enterprise grade service to build and deploy models at scale
12. Demo
13. Training NER Model Using PubMedBert on Azure ML Describe Dataset: name: bc5cdr version: 1 local_path: BC5_data.txt bc5cdr.yml Upload to Azure ML: $ az ml data create -f data_bc5cdr.yml Describe Environment: name: transformers-env version: 1 docker: image: mcr.microsoft.com/ azureml/openmpi3.1.2- cuda10.1-cudnn7-ubuntu18.04 conda_file: file: ./transformers_conda.yml transformers-env.yml channels: - pytorch dependencies: - python=3.8 - pytorch - pip - pip: - transformers transformers_conda.yml $ az ml environment create -f transformers-env.yml
14. Training NER Model Using PubMedBert on Azure ML Describe Experiment: experiment_name: nertrain code: local_path: . command: >- python train.py --data {inputs.corpus} environment: azureml:transformers-env:1 compute: target: azureml:AzMLGPUCompute inputs: corpus: data: azureml:bc5cdr:1 mode: download job.yml Create Compute: $ az ml compute create –n AzMLGPUCompute --size Standard_NC6 --max-node-count 2 Submit Job: $ az ml job create –f job.yml
15. Demo
16. Result • COVID-19 not recognized, because dataset is old • Some other categories would be helpful (pharmacokinetics, biologic fluids, etc.) • Common entities are also needed (quantity, temperature, etc.) Get trained model: $ az ml job download -n $ID --outputs
17. Text Analytics for Health  Can be used through Web API or Container Service  Supports:  Named Entity Recognition (NER)  Relation Extraction  Entity Linking (Ontology Mapping)  Negation Detection
18. Entity Extraction + Entity Linking, Negation Detection
19. Relation Extraction
20. Using Text Analytics for Health Pip Install the Azure TextAnalytics SDK: pip install azure.ai.textanalytics from azure.core.credentials import AzureKeyCredential from azure.ai.textanalytics import TextAnalyticsClient client = TextAnalyticsClient(endpoint=endpoint, credential=AzureKeyCredential(key), api_version="v3.1") Create the client: documents = ["I have not been administered any aspirin, just 300 mg or favipiravir daily."] poller = client.begin_analyze_healthcare_entities(documents) result = poller.result() Do the call:
21. Analysis Result I have not been administered any aspirin, just 300 mg or favipiravir daily. HealthcareEntity(text=300 mg, category=Dosage, subcategory=None, length=6, offset=47, confidence_score=1.0, data_sources=None, related_entities={HealthcareEntity(text=favipiravir, category=MedicationName, subcategory=None, length=11, offset=57, confidence_score=1.0, data_sources=[HealthcareEntityDataSource(entity_id=C1138226, name=UMLS), HealthcareEntityDataSource(entity_id=J05AX27, name=ATC), HealthcareEntityDataSource(entity_id=DB12466, name=DRUGBANK), HealthcareEntityDataSource(entity_id=398131, name=MEDCIN), HealthcareEntityDataSource(entity_id=C462182, name=MSH), HealthcareEntityDataSource(entity_id=C81605, name=NCI), HealthcareEntityDataSource(entity_id=EW5GL2X7E0, name=NCI_FDA)], related_entities={}): 'DosageOfMedication'}) aspirin (C0004057) [MedicationName] 300 mg [Dosage] --DosageOfMedication--> favipiravir (C1138226) [MedicationName] favipiravir (C1138226) [MedicationName] daily [Frequency] --FrequencyOfMedication--> favipiravir (C1138226) [MedicationName]
22. Demo
23. Analyzing CORD Abstracts • All abstracts contained in CSV metadata file • Split 400k papers into chunks of 500 • Id, Title, Journal, Authors, Publication Date • Shuffle by date in order to get representative sample in each chunk • Enrich each json file with text analytics data • Entities, Relations • Parallel processing using Azure ML
24. Parallel Sweep Job in Azure ML CORD Dataset (metadata.csv) Output storage (Database) Azure ML Cluster experiment_name: cog-sweep algorithm: grid type: sweep_job search_space: number: type: choice values: [0, 1] trial: command: >- python process.py --number {search_space.number} --nodes 2 --data {inputs.metacord} inputs: metacord: data: azureml:metacord:1 mode: download max_total_trials: 2 max_concurrent_trials: 2 timeout_minutes: 10000 $ az ml job create –f sweepjob.yml … # Parse command-line df = pd.read_csv(args.data) for i,(id,x) in enumerate(df.iterrows()): if i%args.nodes == args.number: # Process the record # Store the result process.py
25. Results of Text Analytics Processing { "gh690dai": { "id": "gh690dai", "title": "Beef and Pork Marketing Margins and Price Spreads during COVID-19", "authors": "Lusk, Jayson L.; Tonsor, Glynn T.; Schulz, Lee L.", "journal": "Appl Econ Perspect Policy", "abstract": "...", "publish_time": "2020-10-02", "entities": [ { "offset": 0, "length": 16, "text": "COVID-19-related", "category": "Diagnosis", "confidenceScore": 0.79, "isNegated": false },..] "relations": [ { "relationType": "TimeOfTreatment", "bidirectional": false, "source": { "uri": "#/documents/0/entities/15", "text": "previous year", "category": "Time", "isNegated": false, "offset": 704 }, "target": { "uri": "#/documents/0/entities/13", "text": "beef", "category": "TreatmentName", "isNegated": false, "offset": 642 }}]}, …
26. Storing Semi-Structured Data into Cosmos DB Cosmos DB – NoSQL universal solution Querying semi-structured data with SQL-like language Paper Paper Entity Entity Relation Collection … …
27. Cosmos DB & Azure Data Solutions • Real-time access with fast read and write latencies globally, and throughput and consistency all backed by SLAs • Multi-region writes and data distribution to any Azure region with the click of a button. • Independently and elastically scale storage and throughput across any Azure region – even during unpredictable traffic bursts – for unlimited scale worldwide.
28. Cosmos DB SQL Queries Get mentioned dosages of a particular medication and papers they are mentioned in SELECT p.title, r.source.text FROM papers p JOIN r IN p.relations WHERE r.relationType='DosageOfMedication’ AND CONTAINS(r.target.text,'hydro')
29. Demo
30. Further Exploration: Jupyter in Cosmos DB SQL in Cosmos DB is somehow limited Good strategy: make query in Cosmos DB, export to Pandas Dataframe, final exploration in Python Jupyter support is built into Cosmos DB Makes exporting query results to DataFrame easy! %%sql --database CORD --container Papers --output meds SELECT e.text, e.isNegated, p.title, p.publish_time, ARRAY (SELECT VALUE l.id FROM l IN e.links WHERE l.dataSource='UMLS')[0] AS umls_id FROM papers p JOIN e IN p.entities WHERE e.category = 'MedicationName'
31. Visualizations
32. How Medication Strategies Change
33. Term relations
34. Term Relations
35. Terms Co-occurence Treatment Medicine
36. Power BI and No Code / Low Code Data Visualization • Connect to data, including multiple data sources. • Shape the data with queries that build insightful, compelling data models. • Use the data models to create visualizations and reports. • Share your report files for others to leverage, build upon, and share.
37. Demo
38. Exploration: PowerBI
39. Conclusions Text Mining for Medical Texts can be very valuable resource for gaining insights into large text corpus. ❶ ❷ A Range of Microsoft Technologies can be used to effectively make this a reality: • Azure ML for Custom NER training / Parallel Sweep Jobs • Text Analytics for Health to do NER and ontology mapping • Cosmos DB to store and query semi-structured data • Power BI to explore the data interactively to gain insights • Cosmos DB Jupyter Notebooks to do deep dive into the data w/Python
40. Resources • Article: https://soshnikov.com/science/analyzing-medical-papers-with-azure-and-text- analytics-for-health/ • Text Analytics for Health • Azure Machine Learning • Cosmos DB • Power BI • Jupyter Notebooks on Azure Machine Learning • MS LEARN
41. Thank You! http://soshnikov.com