Machine Learning for Understanding Biomedical Publications
1 like320 views
This document discusses machine learning techniques for understanding biomedical publications. It describes multi-label classification approaches for semantic indexing of biomedical literature and modality classification of figures. It also discusses ensemble methods, multi-label learning, and applications to tasks like article screening in systematic reviews and PICO sentence identification.
![EVALUATION
Micro recall used as evaluation measure
Up to 20 items per article
Up to 4 unique Grant Agencies (without Grant ID info) per article
Up to 2 unique Grant Agencies per Grant ID
Sample ground truth for an article
{ "pmid":"17082206", "pmcid":"1634735",
"grantList": [
{"agency":"Wellcome Trust" },
{"grantID":"BB/C51320X/1","agency":"Biotechnology and
Biological Sciences Research Council"}]},
54](https://image.slidesharecdn.com/hscbb-tsoumakas-171013190046/85/Machine-Learning-for-Understanding-Biomedical-Publications-47-320.jpg)
Machine Learning for Understanding Biomedical Publications
- 1. MACHINE LEARNING FOR UNDERSTANDING BIOMEDICAL PUBLICATIONS Grigorios Tsoumakas, School of informatics, Aristotle university of Thessaloniki with: A. Anagnostou, A. Fachantidis, A. Lagopoulos, M. Laliotis, N. Markantonatos, Y. Papanikolaou, I. Vlahavas
- 2. ENSEMBLE METHODS 2 Pedro Domingos. 2012. A few useful things to know about machine learning. Commun. ACM 55 “LEARN MANY MODELS, NOT JUST ONE” Anthony Goldbloom. Kaggle CEO. Oct 2015. “As long as Kaggle has been around, it has almost always been ensembles of decision trees that have won competitions. It used to be random forest that was the big winner, but over the last six months a new algorithm called XGboost has cropped up, and it’s winning practically every competition in the structured data category.”
- 3. MULTI-LABEL LEARNING 3 𝑋1 𝑋2 … 𝑋 𝒑 𝑌1 𝑌2 … 𝑌 𝒒 0.12 1 … 12 0 1 … 1 2.34 9 … -5 1 1 … 0 1.22 3 … 40 1 0 … 1 2.18 2 … 8 ? ? … ? 1.76 7 … 23 ? ? … ? 𝑝 input variables 𝑞 binary output variables 𝑚 training examples unknown instances Binary Relevance (BR) • Learns one binary model per label • Ignores label dependencies
- 4. MULTI-LABEL LEARNING FROM BIOLOGICAL DATA Annotation of proteins with functions FunCat, 6 levels, 492 labels, ~9 on avg. GO, 14 levels, 3997 labels, ~35 on avg. Drug discovery (Johnson & Johnson) 743,336 chemical compounds ~13m chemical structure features (sparse) 5,069 biomolecular targets (e.g. proteins) 4
- 5. OUTLINE 1. Semantic indexing of biomedical literature 2. Article screening in systematic reviews 3. Modality classification of biomedical figures 4. PICO sentence identification 5. Funding information extraction 5
- 6. Literatum is Atypon’s online content hosting and management platform Atypon is home to more than one-third of the world’s English-language professional and scholarly journals — more than any other technology company Atypon’s clients include Elsevier, IEEE, MIT Press, Oxford University Press, … Atypon was acquired by John Wiley & Sons in 2016 for $120,000,000 6
- 7. OUTLINE 1. Semantic indexing of biomedical literature Y. Papanikolaou, G. Tsoumakas, M. Laliotis, N. Markantonatos, I. Vlahavas (2017) Large-Scale Online Semantic Indexing of Biomedical Articles via an Ensemble of Multi-Label Classification Models, Journal of Biomedical Semantics 2. Article screening in systematic reviews 3. Modality classification of biomedical figures 4. PICO sentence identification 5. Funding information extraction 7
- 8. 0 200000 400000 600000 800000 1000000 1200000 1950 1953 1956 1959 1962 1965 1968 1971 1974 1977 1980 1983 1986 1989 1992 1995 1998 2001 2004 2007 2010 2013 x $10 CHALLENGE PubMed abstracts 12,834,585 (20Gb) MeSH terms 27,773, ~13 per abstract on avg. Online test setting 3 phases, 5 weeks per phase MeSH terms for 6k to 10k abstracts requested within 21 hours 8 1 million docs / year ≅ 2,740 docs / day
- 9. LABEL FREQUENCY DISTRIBUTION 9 4.3 million abstracts 213 labels with 1 example 1,680 labels with less than 10 examples 4 labels with more than 1 million examples Label frequency Numberoflabels
- 10. PROGRESS 2013 – 2017 10 0,54 0,56 0,58 0,6 0,62 0,64 0,66 0,68 2013 2014 2015 2016 2017 MicroF-Measure Year AUTH Fudan NLM
- 11. PRE-PROCESSING 11 text of title and abstract parsing tokenization, lowercasing, n-gram extraction 10,876,004 10,699,707 3,950,721 duplicate removal journal filtering tf-idf computation, unit length normalization unigrams and bigrams with >5 frequency final feature vectors last 12,000 withheld for evaluation
- 12. kinetics prognosis liposomes polymerskinetics prognosis liposomes polymers 4 1 2 3 LEARNING 12 Biomedical Document Label Ranker Meta Labeler Number of relevant labels: 2 Output is: {prognosis, liposomes} - Any multi-label learning algorithm that can output a ranking of the labels - We used a linear SVM per label and considered their unthresholded output - Regression or (ordinal) classification using original features or label scores/ranks - We used linear SVM regression based on the original features Tang, L., Rajan, S., Narayanan, V.K. “Large scale multi-label classification via metalabeler”, Proc. 18th Int. Conf. on World Wide Web (WWW '09)
- 13. TIME AND SPACE Hardware 4 10-core processors at 2.26 GHz, 1 Tb RAM and 2.4 Tb storage (6 x 600 Gb SAS 10k disks in RAID 5) Parallel learning/use of binary SVMs With 40 threads training & saving takes 36h With 20 threads loading & prediction takes 45m Serialization Storing the models in 2013 required 406 Gb 10x compression due to sparsity (L1 regularized models)
- 14. ENSEMBLE APPROACHES: CLASSIFIER SELECTION Select the model that improves the corresponding F-measure most Jimeno-Yepes, A., Mork, J.G., Demner-Fushman, D., Aronson, A.R.: A one-size-fits- all indexing method does not exist: Automatic selection based on meta-learning. JCSE 6(2), 151-160 (2012) No good for global non-decomposable evaluation measures, like micro-F Iteratively select the model that improves micro-F Fan, R.E., Lin, C.J.: A study on threshold selection for multi-label classification. Technical report, National Taiwan University (2007) Can we trust the evaluation based on only a few positive samples? 14
- 15. MULE: MULTI-LABEL ENSEMBLE 1. Determine the globally best model ℎ∗ on a validation set Globally best model has been determined on positive samples of all labels 2. Determine for each label which model(s) would lead to an improvement of the global evaluation measure compared to ℎ∗ 3. Compare the differences in predictions of each one of these models against the predictions of ℎ∗ using a McNemar test 4. If the NH is rejected for one or more models, select the one for which the NH has lowest probability, otherwise select ℎ∗ Robustness to uncertainty due to label rarity 15
- 16. EMPIRICAL RESULTS IN BIOASQ 16 Model Micro-F Macro-F Vanilla SVMs 0.5568 0.4789 Weighted SVMs 0.5665 0.5102 MetaLabeler 0.5855 0.5488 Labeled LDA 0.3698 0.3010 Ensemble Micro-F Macro-F Improve F 0.5584 (all) 0.5339 (MetaLabeler, Weighted SVM) Improve Micro-F 0.5867 (all) - MULE 0.5892 (all) 0.5492 (MetaLabeler, Labeled LDA)
- 17. 2017 WORK Improved version of Labeled LDA in both speed and accuracy Y. Papanikolaou, G. Tsoumakas, Subset Labeled LDA for Large-Scale Multi-Label Classification, arXiv:1709.05480 Employing word2vec features More ensembles Stacking-based, frequency-based Deep learning models Deep MLP, CNN 17
- 18. OUTLINE 1. Semantic indexing of biomedical literature 2. Article screening in systematic reviews A. Anagnostou, A. Lagopoulos, G. Tsoumakas, I. Vlahavas (2017) Combining Inter-Review Learning-to-Rank and Intra-Review Incremental Training for Title and Abstract Screening in Systematic Reviews, eHealth Lab of the 8th Conference and Labs of the Evaluation Forum (CLEF) 3. Modality classification of biomedical figures 4. PICO sentence identification 5. Funding information extraction 18
- 19. DIAGNOSTIC TEST ACCURACY (DTA) REVIEWS Title Thromboelastography (TEG) and rotational thromboelastometry (ROTEM) for trauma-induced coagulopathy in adult trauma patients with bleeding Ovid MEDLINE query (Thrombelastogra$ or Thromboelastogra$ or (thromb$ adj2 elastogra$) or TEG or haemoscope or haemonetics).mp Thrombelastography/ (thromboelasto$ or thrombelasto$ or (thromb$ adj2 elastom$) or (rotational adj2 thrombelast) or ROTEM or "tem international").mp. 1 or 2 or 3 exp animals/ not humans.sh. 4 not 5 limit 6 to yr="1970 -Current" 19
- 20. THE DATA Training data 20 topics, retrieved articles and relevance after abstract/content screening Test data 30 topics and retrieved articles 20 1 10 100 1000 10000 100000 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 Pos Neg
- 21. OUR APPROACH Hybrid classification mechanism Inter-topic model, learning title/abstract relevance across different topics Intra-topic model, learning title/abstract relevance within a specific topic
- 22. INTER-TOPIC MODEL Learning-to-rank binary classifier Features assessing the similarity of the document (title, abstract) with the topic (title, ovid query) Common terms Levenshtein distance Cosine similarity BM25
- 23. INTRA-TOPIC MODEL: INITIAL TRAINING 23 TF-IDF
- 24. INTRA-TOPIC MODEL: ITERATIVE TRAINING Parameters Initial size 𝑘 {5, 10} 1st batch size {1} 1st threshold {200, 300} 2nd batch size {50, 100} 2nd threshold {1000, 2000} 24 TF-IDF
- 25. RESULTS Inter-Topic: eXtreme Gradient Boosting (XGBoost) Intra-topic: Support Vector Machine (SVM) 25 k 1st S 1st T 2nd S 2nd T 5 1 200 100 2000 10 1 300 100 2000 10 1 200 100 1000 10 1 200 50 2000 AP Last R@10 R@20 AUR 0.3 2143 0.66 0.87 0.93 0.29 2124 0.66 0.87 0.92 0.29 2183 0.66 0.87 0.92 0.29 2119 0.66 0.87 0.92
- 26. RESULTS 2nd in 14 teams behind University of Waterloo 26
- 27. FUTURE WORK More features, semantic representations Word2vec Glove LDA Under-sampling CLEF eHEALTH 2018 featuring again the same task
- 28. OUTLINE 1. Semantic indexing of biomedical literature 2. Article screening in systematic reviews 3. Modality classification of biomedical figures A. Lagopoulos, A. Fachantidis, G. Tsoumakas (2017), Multi-Label Modality Classification for Figures in Biomedical Literature, 30th IEEE International Symposium on Computer-Based Medical Systems (CBMS) 4. PICO sentence identification 5. Funding information extraction 28
- 29. PUBMED CENTRAL (PMC) More than 4 million figures available Great source of information for biomedical research, education and clinical decision Lack of associated meta-data impede access to this information 29
- 30. 30 different modalities as proposed by ImageCLEF
- 31. SIMPLE VS COMPOUND Simple (60%) Compound (40%)
- 32. THE STANDARD APPROACH Compound Figure Detection Multi-class Model Simple Figure Compound Subfigures Figure Separation Algorithm Figure separation is not perfect (~85%) Figure isolation ⇒ information loss
- 33. OUR APPROACH: MULTI-LABEL CLASSIFICATION No use of figure separation algorithm Three different multi-label learning approaches: • Simple • Standard • Extended
- 34. SIMPLE MULTI-LABEL APPROACH Multi-label model Compound Simple Training Prediction
- 35. STANDARD MULTI-LABEL APPROACH Compound Figure Detection Simple Compound Multi-class Model Multi-label Model Compound Figure Detection Multi-class Model Multi-label Model Compound Simple PredictionTraining
- 36. EXTENDED MULTI-LABEL APPROACH Compound Figure Detection Simple Compound Multi-class Model Multi-label Model Compound Figure Detection Multi-class Model Multi-label Model Compound Simple PredictionTraining
- 37. MODEL TRAINING Feature Extraction from JPEG • BVLC model - Caffe1 • Deep learning (1.2 million images) • 4096 visual features/figure Linear Support Vector Machines (SVMs) • Scikit-learn2 • One-vs-Rest transformation (multiple binaries) 37 1 HTTP://CAFFE.BERKELEYVISION.ORG/ 2 HTTP://SCIKIT-LEARN.ORG/
- 38. IMAGECLEF 2016 DATASET 20.985 Figures 1.568 Compound No simple figures with categories Extracted subfigures as simple Split 40% - 60% (compound – subfigures) in order to follow the distribution of PMC
- 39. RESULTS Approach F1-Macro F1-Micro F1-Samples Standard (100% separation) 0.3569 0.7786 0.7912 Simple multi-label 0.3139 0.7581 0.7215 Standard multi-label 0.3270 0.7667 0.7726 Extended multi-label 0.3309 0.7666 0.7728
- 40. THE SYSTEM Web app Weekly updates from PMC Extended multi-label approach Easy search & filtering by modality Build with Apache Solr & AngularJS Available @ atypon.csd.auth.gr/medieval/
- 41. FUTURE WORK Learning approach Textual representation based on caption and full text Medieval system Crowdsourcing Active learning Gamification 48
- 42. OUTLINE 1. Semantic indexing of biomedical literature 2. Article screening in systematic reviews 3. Modality classification of biomedical figures 4. PICO sentence identification 5. Funding information extraction 49
- 43. 50
- 44. PICO SENTENCE IDENTIFICATION 1st version, ~0.51 F-measure 100 abstracts annotated by computer science PhD students Sentence representation (w2v > tf-idf), structural features MLPs, GaussianNB > SVMs, XGBoost 2nd version 120 more abstracts to be annotated by medical experts Additional feature engineering Semi-supervised learning approaches Deep learning approaches revisited 51
- 45. OUTLINE 1. Semantic indexing of biomedical literature 2. Article screening in systematic reviews 3. Modality classification of biomedical figures 4. PICO sentence identification 5. Funding information extraction 52
- 46. NEW TASK IN BIOASQ 2017 Challenge tasks Full Grant extraction, as combination of Grant ID and Grant Agency Grant ID extraction, regardless of the corresponding Grant Agency Grant Agency extraction, regardless of the specific Grant ID 104 agencies, as considered in the indexing procedure of NLM 53 Timespan Articles Grant IDs Grant Agencies Training set 2005 – 2013 63k 112k 128k Dry run set 2013 – 2015 15k 26k 31k Test set 2015 – 2017 23k - -
- 47. EVALUATION Micro recall used as evaluation measure Up to 20 items per article Up to 4 unique Grant Agencies (without Grant ID info) per article Up to 2 unique Grant Agencies per Grant ID Sample ground truth for an article { "pmid":"17082206", "pmcid":"1634735", "grantList": [ {"agency":"Wellcome Trust" }, {"grantID":"BB/C51320X/1","agency":"Biotechnology and Biological Sciences Research Council"}]}, 54
- 48. RESULTS 55 AUTH: Simple approach based on regular expressions Fudan: Regular expressions combined with machine learning Grant ID Grant Agency Full Grant Fudan 0.9705 0.9907 0.9526 AUTH 0.9498 0.9862 0.9412 DZG 0.9235 0.9122 0.8443 BioASQ 0.8167 0.8312 0.7174
- 49. WRAP UP 1. Semantic indexing of biomedical literature 2. Article screening in systematic reviews 3. Modality classification of biomedical figures 4. PICO sentence identification 5. Funding information extraction 56
- 50. MACHINE LEARNING FOR UNDERSTANDING BIOMEDICAL PUBLICATIONS Grigorios Tsoumakas, School of informatics, Aristotle university of Thessaloniki with: A. Anagnostou, A. Fachantidis, A. Lagopoulos, M. Laliotis, N. Markantonatos, Y. Papanikolaou, I. Vlahavas