Clinical Evidence Extraction from Electronic Health Records and - - PowerPoint PPT Presentation

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature

THE AUSTRALIAN E-HEALTH RESEARCH CENTRE

Hamed Hassanzadeh and Anthony Nguyen 31 July 2018

Problem

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature| Hamed Hassanzadeh 2 |

Looking for relevant information

• Radiology reports
• Discharge summaries
• Progress notes
• Biomedical publications
• Different tasks
• Different clinical evidence

AI Solution

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature| Hamed Hassanzadeh 3 |

Publications and EHRs Publication A Patient Publication B Publication C

EHRs Publications Related to Related to Similar to

Methodology

• Deep Learning: based on Artificial Neural Networks that are

biologically-inspired paradigms

• Enables a computer to learn from observed data

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature| Hamed Hassanzadeh 4 |

Methodology (cont.)

Why Deep Learning?

• Conventional Machine Learning Approaches:

– Task-specific feature engineering – Institution-centric design – Transferability issue

• Deep Learning:

– Numerical representations as input – More generalizable – Transferable

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature| Hamed Hassanzadeh 5 |

Methodology (cont.)

• Word representation

– Word2vec

• Deep Learning

– Convolutional Neural Network

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature| Hamed Hassanzadeh 6 |

Methodology (cont.)

• Convolutional Neural Networks (CNN)

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature| Hamed Hassanzadeh 7 |

the There … is a fracture at distal shaft Target Classes

Vector representations of each word in the document (Embedding Layer) Convolutional layer with multiple filter widths Max-pooling layer Fully connected layer

Data

• Four datasets:
• Benchmark
• Support Vector Machines (SVM)
• Random Forest (RF)

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature| Hamed Hassanzadeh 8 |

DATA SET TYPE OF EVIDENCE TYPE OF DOCUMENTS

i2b2-2010

• Medical Test
• Problem
• Treatment

Clinical documents (progress reports) ShARe/CLEF

• Disorder

Clinical documents (discharge summaries) NICTA-PIBOSO

• Intervention
• Problem/Population
• Outcome

Biomedical publications (abstracts) ED-Radiology

• Abnormality

Clinical Document (Radiology reports – three different hospitals)

Results

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature| Hamed Hassanzadeh 9 |

• Results on public data

– Goal: validate generalisability (multiple tasks)

I2B2-2010 SHARE/CLEF NICTA-PIBOSO TEST TREATMENT PROBLEM DISORDER INTERVENTION POPULATION OUTCOME

RF

0.59 0.65 0.73 0.74 0.02 0.0 0.51

SVM

0.80 0.84 0.85 0.85 0.0 0.21 0.68

Our Approach

0.90 0.93 0.93 0.87 0.41 0.57 0.75

Results (cont.)

• Results on ED-Radiology

– Goal: validate transferability (multiple institutions)

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature| Hamed Hassanzadeh 10 |

RBWH RCH GCH RF 0.80 0.83 0.84 SVM 0.83 0.91 0.91 Our approach 0.91 0.94 0.94

Conclusion

• A generalizable and transferable solution

– Different clinical document (progress notes, discharge summaries) – Biomedical publications – Different institutions

• Future work:

– Expand the solution and validate it over more and bigger clinical problems – Translating the outcome more into practice

Clinical Evidence Extraction from Electronic Health Records and Biomedical Literature| Hamed Hassanzadeh 11 |

Hamed Hassanzadeh, Postdoctoral Fellow e hamed.hassanzadeh@csiro.au

Thank you

HEALTH & BIOSECURITY