Transcript Real time risk prediction in healthcare

Avionics EUROPE 2012
The context in Aviation Safety
Flight safety improvement has been challenged during several decades:
● The number of departures has been steadily increasing
● Aviation accident rate has been significantly reduced
● Air traffic is projected to grow in the future
An increase in the number of accidents is
not acceptable.
How to reduce the accident rate further ?
2
3
Key challenge #1 for enhancing current FDM programmes
● A huge increase in available flight data:
● Processing power & digital storage increases every year
e.g. a typical acquisition will produce:
~10 Mbytes / aircraft / day
~2000 parameters recorded real time
~100 flight reports can be monitored
● EU-OPS/CAA guidance:
● ~60 events to be monitored
● A lot of FDM data available, but:
● Only a small part is fully exploited,
● Even on the most modern aircraft
4
Key challenge #2 for enhancing current FDM programmes
● Current FDM practices: an event-based strategy
● Predefined rules & event sets
● Highly optimized by each airline
● Rather reactive than proactive
● Unable to find all
singularities in
flight data
● Any information stored in
flight data but never used ?
Actual weather during landing :
no visibility !
Airbus A340-313, 02/08/2005 - Toronto.
Most of the aircraft was destroyed. This situation cannot be detected with the
traditional approach of existing FDA systems.
A long landing
under very bad
weather conditions
5
Key challenge #3 for enhancing current FDM programmes
● Enhance automation of current FDM systems, they mostly rely on human expertise for all tasks:
●
●
●
●
●
●
Event invalidation
Statistics / trends publication
Detection of causal factors of incidents
Identification of common causal factors
Understanding the relation with external data
….
● A heavy workload:
● e.g. a B737 operator (10 A/C), a typical set of 80 events monitored
● 4.5% of class 3 events
● 15 000 flights analysed per year
If event set is increased, workload is increased
6
SVETLANA – The Project
● The SVETLANA project is an European Community – Russia cooperation in the field of aviation safety and
maintenance improvement.
● The project is aimed at improving the capabilities of flight data monitoring programmes for civil
aviation.
● SVETLANA is a 2 years project leaded by SAGEM, started in October 2010 in the frame of the European
FP7 Work Programme.
● An End User board is associated to SVETLANA:
● Composed by experts from EU & Russian airlines, Regulatory Organizations, Standardisation Groups
● Provide consultancy support to assess the operational process, quantify operational flight safety benefit, assess enhanced
maintenance benefit
7
Key challenge #4 for enhancing current FDM programmes
● A poor link to external data:
●
●
●
●
●
●
Weather data
Pilot report data
Other Safety Management Systems
Other airline databases
Additional information from the aircraft
…
● Several other challenges prevent a thorough
analysis of 100% of all the available flight data:
● the lack of standards for a common FDA methodology,
● the lack of links with other data sources, the lack of regulations to share data,
● the power of unions, the lack of knowledge to use flight data for other purposes, …
8
SVETLANA – The Consortium
●
Sagem Défense Sécurité
Project coordination, expert in FDM and Information Systems for airlines
●
Stichting Nationaal Lucht – En Ruimtevaartlaboratorium
Expert in the field of aviation safety
●
Lancaster University
Specialized in data mining and extraction of knowledge from data streams
●
JSC United Aircraft Corporation
Aircraft manufacturer
●
JSC Concern Avionica
Expert in information monitoring (Flight Data Recorder)
●
FSUE Institute of Aircraft Equipment
Specialized in Russian civil aircraft equipment design
●
ALMA Consulting Group SAS
Project management, communication and dissemination
9
SVETLANA – Objectives
The main objectives of SVETLANA are:
● to offer a systematic enhanced flight data analysis process
● to propose SVETLANA as a standardized data analysis process
● on existing fleet / on existing aircraft
● for operational safety purpose
● additional benefit expected for enhanced maintenance
● to enhance analysis capabilities & services in existing and future FDM systems
The approach used in SVETLANA:
● A two steps cycle:
● Use of automated flight data processing tools
● Human expertise for validation and decision making process
● Knowledge extraction based on Data Mining techniques
● Advanced algorithms to identify fault signature and predict abnormal behaviour
● A feedback loop for shorter analysis and detection cycles
10
SVETLANA – answers to enhanced FDM challenges
● Existing systems are model based:
● physics based, e.g. aeronautics rules
● Knowledge based rules, e.g. expert systems,
Increasing number of parameters lead to more complex models, difficult to make prognostics and to adapt to regulation evolutions
● SVETLANA will add full data driven methods: learning & on-line capabilities
SVETLANA
11
SVETLANA – answers to enhanced FDM challenges
● SVETLANA Data Mining algorithms sample result of anomaly detection (red plot):
● Compared to current FDM systems (blue plot)
Detection of singularities without predefined event specification – Source: Lancaster University, research is supported by previous published work results:
- P. Angelov, P. Sadeghi-Tehran, R. Ramezani, A Real-time Approach to Autonomous Novelty Detection and Object Tracking in Video Stream, International Journal of Intelligent Systems, vol.26 (3) 189-205, March 2011,
- P. Angelov, R. Ramezani, X. Zhou, Autonomous Novelty Detection and Object Tracking in Video Streams using Evolving Clustering and Takagi-Sugeno type Neuro-Fuzzy System, 2008 IEEE International Joint Conference on Neural Networks within the IEEE
World Congress on Computational Intelligence, Hong Kong, June 1-6, 2008, pp.1457-1464, ISBN 978-1-4244-1821-3/08,
- P. Angelov, Machine Learning (Collaborative Systems) patent (WO2008053161, priority date: 1 Nov. 2006; international filing date 23 Oct. 2007); USA publication 11 Feb. 2010, # 2010-0036780
12
SVETLANA – answers to enhanced FDM challenges
● Improvements on existing FDA systems:
● Capability to routinely process a large amount of data
● Learning capability: no need to process again the complete database
● Advanced clustering techniques
● A better automation from higher
aircraft independence:
● Detection independent from AC type
● Detection independent from the number
and nature of parameters
● Independent of airline operating policy
● Capability to handle recording failures
● Detection operates on a single flight compared
to many, or many flights compared to many
THEORY
REALITY
Sample clustering of numerous flight data – Source: SAGEM
13
SVETLANA – answers to enhanced FDM challenges
● Improvements on existing FDA systems:
● Robust flight phase detection, handling more recording failures
Jet engine (Boeing 737)
Turboprop engine (Dash8 Q400)
Current FDA system
SVETLANA
Robust and aircraft independent flight phase identification – Source: SAGEM
14
SVETLANA – answers to enhanced FDM challenges
●
More automation:
Split flight in phases
1. Building a chain of automated analysis steps:
● Self learning algorithms
2. Training the system with:
Feature selection
3. Enhancing aircraft “independence” in FD analysis
Anomaly detection
● Flight data from real operations
● Simulated data containing known issues
Parameter identification*
* Identification of parameter responsible of anomaly detection
Prognostics
15
SVETLANA – Expected results and next steps
● Expected benefits:
●
●
●
●
●
Flight safety improvement
Maintenance support improvement
Enable more analysis with same or lower human intervention
Provide feedback to stakeholders and improve FDM process
Enhance flight data analysis
● Exploitation of the results:
● Integration in future versions of
SAGEM’s AGS
Analysis Ground Station (AGS) – Source: SAGEM
16
Hospital
Online
Uncertainty
and Stability
Estimation
IDENTIFYING ‘TIPPING POINTS’ IN
CRITICALLY ILL PATIENTS
Background
25,000 patients/year cardiac surgery – 3% mortality
4,000 patients/year elective AAA repair – 3% mortality
3,500 patients/day in intensive care – 5% mortality
Aviation Vs Healthcare
Aviation Vs Healthcare
Aviation Vs Healthcare
WHO surgical safety checklist
Adverse event rate fell from
1.5% to 0.8% (p=0.003)
ICU monitoring
Heart rate, blood pressure,
respiratory function, arterial and
venous pressure, urine output
Threshold alarms
≈450 data points/patient/day
≈700,000 data points/patient/day
Individualised ‘tipping points’
Significant physiological variation between patients
Early identification of patient deterioration and prompt treatment
improves outcomes
Often subtle signs of early deterioration are missed
Objective
To develop a real time risk prediction system for
the early identification of adverse events in
intensive care
Modelling
Patient factors
Outcomes
Co-morbidity
Treatment escalation
ECG
Blood pressure
Surgery outcomes
Respiratory function
Complications
Mortality
Urine output
Blood tests
Patient stability
Patient Transfer Cycle
Normal Cycle
ICU
Theatre
Home
Ward
Minimize relapse paths
Electrocardiography
QRS Complex
R
PR
Segment
ST
Segment
T
P
Q
PR Interval
S
QT Interval
MIMIC II database
Managed by MIT
Contains clinical and waveform data
Single hospital
ICU patients 2001 - 2008
Monitoring ECGs
Detect and measure abnormal rhythms
Expected to have constant period
Changes in amplitude, period and polarity of a specific
wave may help diagnose certain pathologies
The accuracy and interpretation of the signal varies
according to the placement of the leads
Modelling ECGs
Waves are expected to have constant periods and
amplitudes
Amplitudes and time intervals are treated as random
variables for each lead and a stability score is assigned
according to the regularity of the signal
If an unlikely arrival is detected in more than one lead, then
the stability score is affected
The score can be used to create warning systems for doctors
and nurses that don’t depend only on thresholds and that
can be adjusted to an individual patient without much data
Example: R wave arrivals
Future work
Incorporation of additional physiological parameters into
overall stability model
Development of real time clinical event prediction model
Initiation of clinical effectiveness trials