20
| | ETH Zurich, Data Analytics Lab Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR Carsten Eickhoff

Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

  • Upload
    others

  • View
    5

  • Download
    0

Embed Size (px)

Citation preview

Page 1: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab

Lorenz Kuhn & Carsten Eickhoff

21.09.2016 1

Implicit Negative Feedback in Clinical IR

Carsten Eickhoff

Page 2: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 2

Medical Literature: Scale & Growth

21.09.2016Carsten Eickhoff

▪ 25 Million articles indexed in MedLine▪ Recently growing by ~1 Million new articles per year▪ Manual search and processing become infeasible▪ Search technology is needed▪ Goal: Processing the collection in < 20 mS

Page 3: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 3

Patient-centric Search Scenarios

21.09.2016Carsten Eickhoff

Page 4: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 4

Patient-centric Search Scenarios

21.09.2016Carsten Eickhoff

Page 5: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 5

Patient-centric Search Scenarios

21.09.2016Carsten Eickhoff

Page 6: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 6

Patient-centric Search Scenarios

21.09.2016Carsten Eickhoff

Page 7: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 7

Patient-centric Search Scenarios

21.09.2016Carsten Eickhoff

Page 8: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 8

Medical IR

A 58-year-old African-American woman presents to the ER with episodic pressing/burning chest pain that began two days earlier for the first time in her life. The pain started while she was walking, radiates to the back, and is accompanied by nausea, diaphoresis and mild dyspnea, but is not increased on inspiration. The latest episode of pain ended half an hour prior to her arrival. She is known to have hypertension and obesity. She denies smoking, diabetes, hypercholesterolemia, or a family history of heart disease. She currently takes no medications. Physical examination is normal. The EKG shows nonspecific changes.

21.09.2016Carsten Eickhoff

Page 9: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 9

Medical IR

A 58-year-old African-American woman presents to the ER with episodic pressing/burning chest pain that began two days earlier for the first time in her life. The pain started while she was walking, radiates to the back, and is accompanied by nausea, diaphoresis and mild dyspnea, but is not increased on inspiration. The latest episode of pain ended half an hour prior to her arrival. She is known to have hypertension and obesity. She denies smoking, diabetes, hypercholesterolemia, or a family history of heart disease. She currently takes no medications. Physical examination is normal. The EKG shows nonspecific changes.

21.09.2016Carsten Eickhoff

Page 10: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 10

Medical IR

A 58-year-old African-American woman presents to the ER with episodic pressing/burning chest pain that began two days earlier for the first time in her life. The pain started while she was walking, radiates to the back, and is accompanied by nausea, diaphoresis and mild dyspnea, but is not increased on inspiration. The latest episode of pain ended half an hour prior to her arrival. She is known to have hypertension and obesity. She denies smoking, diabetes, hypercholesterolemia, or a family history of heart disease. She currently takes no medications. Physical examination is normal. The EKG shows nonspecific changes.

21.09.2016Carsten Eickhoff

Page 11: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 11

Medical IR

A 58-year-old African-American woman presents to the ER with episodic pressing/burning chest pain that began two days earlier for the first time in her life. The pain started while she was walking, radiates to the back, and is accompanied by nausea, diaphoresis and mild dyspnea, but is not increased on inspiration. The latest episode of pain ended half an hour prior to her arrival. She is known to have hypertension and obesity. She denies smoking, diabetes, hypercholesterolemia, or a family history of heart disease. She currently takes no medications. Physical examination is normal. The EKG shows nonspecific changes.

21.09.2016Carsten Eickhoff

▪ Negated phrases contain valuable information

Page 12: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 12

Detrimental Effects on Search Performance

21.09.2016Carsten Eickhoff

Page 13: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 13

Medical IR

A 58-year-old African-American woman presents to the ER with episodic pressing/burning chest pain that began two days earlier for the first time in her life. The pain started while she was walking, radiates to the back, and is accompanied by nausea, diaphoresis and mild dyspnea, but is not increased on inspiration. The latest episode of pain ended half an hour prior to her arrival. She is known to have hypertension and obesity. She denies smoking, diabetes, hypercholesterolemia, or a family history of heart disease. She currently takes no medications. Physical examination is normal. The EKG shows nonspecific changes.

21.09.2016Carsten Eickhoff

▪ Negated phrases contain valuable information▪ Using them as query terms can be problematic

Page 14: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 14

Medical IR

A 58-year-old African-American woman presents to the ER with episodic pressing/burning chest pain that began two days earlier for the first time in her life. The pain started while she was walking, radiates to the back, and is accompanied by nausea, diaphoresis and mild dyspnea, but is not increased on inspiration. The latest episode of pain ended half an hour prior to her arrival. She is known to have hypertension and obesity. She denies smoking, diabetes, hypercholesterolemia, or a family history of heart disease. She currently takes no medications. Physical examination is normal. The EKG shows nonspecific changes.

21.09.2016Carsten Eickhoff

▪ Negated phrases contain valuable information▪ Using them as query terms can be problematic▪ Naive approach: Remove all negated phrases

Page 15: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 15

Medical IR

A 58-year-old African-American woman presents to the ER with episodic pressing/burning chest pain that began two days earlier for the first time in her life. The pain started while she was walking, radiates to the back, and is accompanied by nausea, diaphoresis and mild dyspnea, but is not increased on inspiration. The latest episode of pain ended half an hour prior to her arrival. She is known to have hypertension and obesity. She denies smoking, diabetes, hypercholesterolemia, or a family history of heart disease. She currently takes no medications. Physical examination is normal. The EKG shows nonspecific changes.

21.09.2016Carsten Eickhoff

▪ Negated phrases contain valuable information▪ Using them as query terms can be problematic▪ Naive approach: Remove all negated phrases▪ Alternative: Use them as negative feedback

Page 16: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 16

Formal Patient-centric Retrieval

21.09.2016Carsten Eickhoff

Standard Retrieval Model

Negation Filtering

Negative Feedback

Page 17: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 17

Experimental Setup

21.09.2016Carsten Eickhoff

▪ TREC Clinical Decision Support Track▪ 30 anonymized clinical notes▪ 733,000 PubMed articles▪ Find relevant literature for each patient▪ Evaluated by NLM experts

Page 18: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 18

Results

21.09.2016Carsten Eickhoff

Page 19: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 19

Conclusion

21.09.2016Carsten Eickhoff

▪ Clinical data science becomes crucial as collections scale

▪ Accounting for phrase compositionality is important

▪ E.g., negated phrases

▪ Negations detriment state-of-the-art retrieval methods

▪ Explicitly modelling and using this type of information helps

▪ Future (ongoing) work: Deep neural networks for modelling phrase compositionality

Page 20: Implicit Negative Feedback in Clinical IR · Lorenz Kuhn & Carsten Eickhoff 21.09.2016 1 Implicit Negative Feedback in Clinical IR |ETH Zurich, Data Analytics Lab | 2 Medical Literature:

||ETH Zurich, Data Analytics Lab 20

Thank You!

21.09.2016Carsten Eickhoff