Towards identifying emergency patients with unconcerning brain - - PowerPoint PPT Presentation

The University of Sydney Page 1

Towards identifying emergency patients with unconcerning brain computed tomography scans

Presented by Dr Aldo Saavedra Dr Madhura Killedar

• n behalf of

Prof Jonathan Morris Dr Joel Nothman Seven Guney Dr Catherine Naidoo Dr Lucy Blumer Peter Thiem Dr Felicity Gallimore

The University of Sydney Page 2

Introduction

• Overall aims:
• Harness the information

captured by the eMR

• Identify problems that

can be solved with the data

• Multi-disciplinary team from

Sydney University and Royal North Shore Hospital (RNS).

• RNS is a tertiary teaching

hospital of the University of Sydney

• It services 17% of the

population in NSW

• Increasing number of presentations to the

emergency department (ED) are placing stress on an already busy health care system (4% between 2017 and 2018)

The University of Sydney Page 3

Computed Tomography Scan – CT Scans

• Focus on a resource intensive

investigation: CT scans

• At RNS there are 3 scanners with a

daytime team 4-6 consultant radiologist and 6-10 registrars

• One registrar after 8:pm
• The actual scan can take 5-15 minute.
• It can take 3-4 staff to transfer an

elderly patient on/off the bed – 10 minutes.

• A few minutes for write a report if

radiologist is not interrupted.

University of Virginia

The University of Sydney Page 4

Motivation

The focus

• f our

study

An increase of ≈ 13 per quarter

The number of scans in Q1 of 2013 was 387

• 69 different

types of scans performed by ED in three years

• CT Brain scans

correspond to ≈ 50%

• f all ED Scans

The University of Sydney Page 5

Project goals using data

Data subset: EMRs for 5600 encounters from RNS emergency between 2013 & 2016 in which CT brain scans were taken for patients >16 yrs old – are CT scan requests increasing with time? – do these CT scans have positive/negative findings? – can we classify findings based on text from the CT scan report? – is the information captured by the electronic medical record (eMR) during an emergency presentation enough to predict whether the scan will have a negative finding?

The University of Sydney Page 6

Determining negative findings from CT scan report text

– Two radiology registrars labelled 843 reports manually

– patient demographics made available but not needed – conclusion text was sufficient – high level of agreement

– Two classes of findings

– Negative if both registrars agree negative

• r expected for a given age

– Positive: otherwise – Remain conservative

CT Brain performed on 29-MAY-2015 at 11:53 AM Dictated By: Dr X. Xxxxx Typed By: Dr X. Xxxxxx Approved By: Dr Yyyyy Yyyyyyy Clinical History: Unwitnessed fall at NH; on anticoagulants - clopidogrel. Findings: No acute intracranial haemorrhage. A lacunar infarct is noted in the left basal ganglia (likely chronic) unaltered since 23/05/2015. Bilateral periventricular white matter hypodensity is unaltered since the previous study compatible with chronic microvascular ischaemic

• change. Grey-white matter differentiation is otherwise
• maintained. The ventricles and CSF spaces are enlarged in

keeping with volume loss. Patchy mucosal thickening involving the visualised paranasal

• sinuses. No fracture of the skull vault or visualised facial

bones. Conclusion: No acute intracranial haemorrhage.

The University of Sydney Page 7

Classifying CT scan findings

• 2a. Rule-based system assesses sentences,

eliminates if describing normality or absence of finding, looks for any remaining

• very high sensitivity
• 2b. Machine learning system used bag-of-

words features

• more balanced sensitivity and specificity

No acute intracranial pathology identified. No space-occupying lesion identified. No acute intracranial pathology identified. No space-occupying lesion identified.

• 3. Hybrid machine learning classifier (used rule-based prediction as ML feature)

– trained against manual classification – very high sensitivity (93%) , but improved specificity (96%)

• 1. Extract conclusion text

The University of Sydney Page 8

Increasing Negative Findings?

77%

5% 18%

14% 5% 81%

83%

14% 3%

80% 76% 81% 80% 80% 81% 81% 77% 82%

The University of Sydney Page 9

Information from EMR

Demographics – age, gender, indigenous status, preferred language Vital Signs (on arrival) – blood pressure, SpO2, respiratory rate Pathology test results – Haemoglobin, Platelet Count, White Cell Count, C-Reactive Protein, Creatinine, AST, PT, APTT, INR, Glucose (Rand) Emergency Triage Form (Presenting Problem) – headache, falls, injury, head, dizziness, pain, limb, altered, syncope, sensation, weakness, faint, seizure, eye, mh, abnormal, vision, loc, care, speech, difficulties, chest, facial Emergency Triage Form (Presenting Information for Tracking Board) – pain, headache, head, fall, loc, alert, biba, well, neck, onset, nausea, limb, dizziness, chest, facial, side, perfused, hit, equal, weakness, limbs, vision, sided, back, arm, speech, droop

• Clinical consultation to determine relevant fields
• Raw data categorized relative to medical standard reference values
• Ensure data is recorded before CT scan
• Mutual information points at redundancies

The University of Sydney Page 10

Cohorts of patient encounters

Distinguishing feature: presenting problem

Cohort Cohort Cohort Cohort

The University of Sydney Page 11

Presenting problem Number of encounters # encounters with positive findings % encounters with positive findings

Headache

999 129 13%

Falls

866 86 10%

Injury

719 74 10%

Dizziness

450 46 10%

Limb

283 36 13%

Syncope

244 30 12%

Weakness

221 35 16%

Pain

290 30 10%

Altered

257 26 10%

Seizure

172 20 12%

Eye

154 17 11%

Mental Health

133 8 6%

Most common presenting problems

The University of Sydney Page 12

Learning and Predicting

– Predict +ve/-ve CT scan findings via a percentage probability of +ve – Bayesian Generalized Linear Model

– performs logistic regression – each of EMR fields are features – trained against CT scan report classification

– Bayesian framework accounts for uncertainties, and model removal of features – Calibrate separately for each presenting problem: headache, falls, injury – Learns from 80% of data (2400 encounters) – Predicts for each of remaining 20% (600 encounters)

The University of Sydney Page 13

Evaluating the machine

Set conservative threshold: 2% of positive findings fall below 28 out of 482 (6%) negative findings are identified Corresponds to 300 encounters from total dataset of 5600

Likely negative Likely positive

Combined 600 results for encounters: headaches, falls, injury

The University of Sydney Page 14

Summary

– Increase in brain CT scans at RNS – Large value in electronic medical records

– machine learning on report text to classify CT scan findings – characteristics of encounters – identifying cohorts – predict findings for patient encounters

– Potential to increase sensitivity:

–

• ther models

– incorporate previous studies – varied diagnoses – additional fields in EMR