Improving care of people with mental health

farmpaintlickInternet και Εφαρμογές Web

21 Οκτ 2013 (πριν από 3 χρόνια και 5 μήνες)

80 εμφανίσεις

Improving care of people with mental health
problems using the
Galatean

Risk and Safety Tool
(
GRiST
)

Christopher
Buckingham,
Computer
Science, Aston University

Ann
Adams,
Medical
School, University of Warwick

April 10
th
, 2013

The potential for IAPT services


www.egrist.org

LUFC

Elland

Road

Risks associated with mental health
problems


Suicide


Self harm


Harm to others and damage to property


Self neglect


Vulnerability


Risk to dependents

Our research is about better understanding,
detection, and management

It is aimed at both clinicians and service users

It feeds into the GRiST clinical tool and improved
services

Some of the Research Team

Ann Adams,

& Christopher Mace

University of Warwick

Christopher Buckingham,

Ashish

Kumar, Abu Ahmed

University of Aston

Evidence about

mental
-
health risks

Risk

independent cues

Risk

cue clusters

Risk

cue interactions

specific cue values

occurring together

particular cue

combinations

We know quite a lot

We know a little

We hardly know anything

No explicit integration

RISK

ASSESSMENT

Risk tool

Clinical judgement

Need to connect the information sources

RISK

ASSESSMENT

Risk tool

Clinical judgement

Data hard to extract

Electronic documents: little structure, information buried

Yes, this really is an NHS decision support document

Data not shared

RISK

ASSESSMENT

RISK

ASSESSMENT

Mon

Tue

Fri

RISK

ASSESSMENT

or exploit

the semantic

web

The solution: GRiST


Explicitly models structured clinical judgements


Underpinned by a database with sophisticated statistical
and pattern recognition tools.


linked with empirical evidence


Developed from the start to exploit the semantic web


universally available


ordinary web browsers


Designed as an interactive tool with sophisticated
interface functionality


Provides a common risk language with multiple
interfaces


collecting information


providing advice


Supports shared decision making and self
-
assessment



The solution:
GRiST


Versions for different populations


older, working age, child and adolescent


specialist services (e.g. learning disability, forensic)


A
whole (health and social care) system approach to risk
assessment


www.egrist.org

Eliciting expertise

Knowledge bottleneck


Extracting expertise


Representational language experts understand


Gain agreement between multiple experts


Lowest common denominator ……

Unstructured Interview


What factors would you consider important to
evaluate in an assessment of someone presenting
with mental health difficulties?


prompts or probes to explore further


46 multidisciplinary mental
-
health practitioners

Mind map with total numbers of experts

results of integrating interview data

12
experts



identifies relevant service
-
user data


“tree” relates data
to risk concepts and top
-
level
risks


information profile for service user



Interview
transcripts

Qs & layers

XSLT

Different risk

screening

tools for

varying

circumstances

and assessors

Lisp or XSLT

Mind map

Tree for pruning

Pruned tree

Data gathering tree

Data gathering tree

with questions and layers

that organise question priority

Fully annotated

pruned tree

mark up

XSLT

All trees are implemented as XML

Multiple populations handled by
instructions in the tree


Work on specifying
different models done
by XML attributes


End
-
users access their
own simple tree


What is XML?

<family>

<brother “john”/>

<sister “
mary
”/>

<daddy “long legs”/>

</family>


Arboreal sculpture


Complete “universal” tree: multiple
overlays

working age

Complete “universal” tree: multiple
overlays

CAMHS

Complete “universal” tree: multiple
overlays

Older

Adults

Complete “universal” tree: multiple
overlays

S
ervice

users

Complete “universal” tree: multiple
overlays

Carers

Complete “universal” tree: multiple
overlays

Friends

Multiple services


Same idea as populations


Customise service requirements


Difference is that they cover all populations


Services so far:


IAPT


Primary Care


Forensic

How not to design and develop


Must be able to meet end
-
user’s changing and
varied requirements

Iterative development for implementing
research results into evolving GRiST and
myGRiST

Agile software

engineering



IAPT demo

If the person says yes

IAPT version

of Grist

just 6 screening

questions

Opens up four subsidiary questions
for IAPT

If the person says yes

Two more IAPT questions are
asked.

Comments and management
information can be added to any
questions

An overall risk judgement is made
along with supporting comments
and risk management information

Risk reports are
generated immediately
and can be downloaded
as a
pdf
.

This shows a summary
just

for suicide

Each risk has a detailed
information profile that explains
where the risk judgement came
from.

comment

action/intervention

gold padlock

silver padlock

red means filled

Interface functionality

Manage patient assessments


Service audit data (
i
)


Service audit data (ii)


myGRiST

myGRiST

Communication


GRiST

Cloud


common data

PHQ
-
9
et al

GAD
-
7

Data sharing

Data exchange

Data integration

Patient
-
centric web of care

clinical perspective

Risk

clinical/service user

Safety

service user

Well
-

being

Current GRiST database (now twice as
big)


96,040 cases of patient data linked to clinical
risk judgements


Different risks


Different age ranges


Precise quantitative input linked with
qualitative free text

f(data)

How we do it

Transparent

Knowledge and reasoning can be understood



Black box



Can’t see how
answer derived

input data

Risk data

output judgement

Risk evaluation

GRiST

cognitive model

Clear explanation for risk judgement

Identifies important risk concepts

Informs interventions

Mathematical models

Optimal prediction of judgement

Validation of cognitive model

Evidence base for cues and relationship with risks

RBFN

BBN

neural net

PCA


secure

trusted

risks

GRiST captures consensus


Preliminary (crude analysis) results for clinical tool


Correlation > 0.8, R
2

= 0.69


87% of 4000 predictions within 1 of the
expert on 11
-
point
scale


No difference if inputs are raw values or
membership grades


So we can model evaluations for different types of user



Clinical Decision Support for Mental Health

www.eGRiST.org



Galatean

Risk Screening Tool

Results

Absolute Error in
Predicting Judgement

87% of predictions have
an error of < +/
-

1

eg

If judgement = 3,

2 < prediction < 4

Less than 3% have an
error of greater than +/
-

2

less than 2

87%

>+/
-

1

13%

less than 2
More than 2
No

Risk

Low

Risk

Medium

Risk

High

Risk

Max

Risk

0

to

2

2

to

4

4

to

6

6

to

8

8

to

10

www.egrist.org