Transcript Estimating recurrence and survival outcomes in cancer using routine

Using routine data
to measure recurrence
in Head and Neck Cancer
Zi Wei Liu
Matt Williams
Adam Gibson
Kate Ricketts
Heather Fitzke
[email protected]
Imperial E-oncology Conference 2015
Defining the problem

Head and neck cancer
–
~6000 new diagnoses of head and neck cancer a
year
–
Strongly related to smoking
–
Increase in incidence recently due to HPV related
H+N cancer
–
~60% present at an advanced stage and require
multi-modality treatment-surgery, radiotherapy,
chemo.
Defining the problem


Recurrence rates in H&N cancer are important

For staff (efficacy)

For patients (prognosis)

For service planning (costs)
Not well measured in routine care population
Relies on patchy manual data entry (9th DAHNO 12%
reported)
What is 'routine data'?


Nationally collected patient data

Uniform coding scheme

Some linked to payments for activity

mandatory data collection
Examples:

HES (hospital episodes statistic)

SACT (Systemic anti-cancer therapy)

RTDS (radiotherapy database)

DBS(Demographic batch service)

Cancer registry data
Hospital episodes statistic

Patient demographics

Inpatient (and now outpatient) attendances

Diagnosis & Procedures

Co-morbidities
SACT & RTDS

SACT & RTDS cancer databases have a minimum
dataset which usually contains the following:

Patient demographics: e.g NHS number, DOB, post
code, consultant code

Primary diagnosis: ICD-10 code, staging, morphology

Regimen, intention of treatment, height and weight,
PS

Start and end date of treatment, intended and actual
treatment delivered

Date of death
Aims and importance of our study



Can we determine recurrence rates and survival times
from routine data ?
How closely do they match manually-measured rates &
times ?
Pilot study
assess feasibility and possible problems

Follow-up study
larger sample size, problems with scaling
Methods



Pilot study:
20 patients with head and neck identified from local MDT
lists

Received radical treatment

Weighted towards those diagnosed at UCH

Weighted towards advanced disease
Paired datasets generated-'manual' and 'routine'
Tests of correlation performed on key clinical outcome
indicators such as overall survival, progression survival
and recurrence events.
Ref: Liu ZW, Fitzke H, Williams M. Using routine data to estimate survival and
recurrence in head and neck cancer: our preliminary experience in twenty
patients. (2013) Clinical Otolaryngology, 38(4):334-9.
Methods

Second expanded study 122 patients

Paired datasets generated-'manual' and 'routine'


Optimization strategies including backdating, time interval
optimization
Survival curves
Ref: Ricketts K, Williams M, Liu ZW, Gibson A. (2014). Automated estimation
of disease recurrence in head and neck cancer using routine healthcare data.
Computer Methods and Programs in Biomedicine. 7(3):412-24.
Methods
Methods
Methods

Date & Site of first head and neck cancer diagnosis code

Radical treatment

Collect HES, RTDS and SACT data (incl. Dates)

If further major surgical resection or palliative
chemotherapy, or palliative RT, assume recurrence

No intention on RT, so used a 3/12 cut-off for
differentiating adjuvant vs. radical salvage RT
Results
 Pilot study:
 20 patients
 13 male
 9 primary oropharynx
 15 LAHNSCC
 Median OS 24.4 months
 Median PFS 9.6 months
Results
 Follow-up Study:
 122 patients
82% locally advanced disease
 51 oropharynx
 26 larynx
 Median OS 88% (1 year), 77% (2 years)
 Median PFS 75% (1 year), 66% (2 years)
Results
 Optimization strategies
– Backdating
– Optimizing time intervals between primary and
secondary treatment
Results
Conditi
ons
No.
patien
ts out
of
bound
s for
routin
e OS
No.
patie
nts
out of
boun
ds for
routin
e PFS
Diagnosis
dates in
agreement
{n = 122}
±1 week /
±1 month
Recurrenc
e dates in
agreement
{n = 40}
±1 week /
±1 month
No. of
recurrenc
e events
correctly
identified
No. of
No. of
recurre recurren
nce
ce
events
events
falsely missed
identifi
ed
Initial
approac
h
7
25
1 week (62)
1 month (97)
1 week (1)
1 month (4)
21
5
19
Backdat
ing
alone
3
23
1 week (61)
1 month
(101)
1 week (5)
1 month (7)
21
5
19
Backdat
ing +
optimis
ed time
interval
s
3
21
1 week (61)
1 month
(102)
1 week (7)
1 month (9)
21
2
19
Results
Results
Results
Pilot study (n=20)
Follow up study (n=122)
OS
95% good agreement
98% good agreement
PFS
80% good agreement
82% good agreement
Recurrence events
10/11 correctly identified 21/40 correctly identified
Discussion




Selected sample

LAHNSCC

Radical treatment only
Reasonable agreement between routine and manual data
Used national-level data, possible to automate, adds to
existing knowledge
Potentially inaccurate, esp. in palliative patients
Discussion

Further optimisation work
HES density data looking at ratio of inpatient to
outpatient attendances to predict recurrence


Measurement of non-OS outcomes
–
In addition to recurrence:
–
PEG dependency rates
–
Tracheostomy dependency rates
Future directions

Phase III study using national cancer data under way

Develop software to automate data handling and analysis


Experiments to optimise algorithm and utilise
modelling to improve accuracy of predictions

Incorporate registry data

First comprehensive automated analysis of national
cancer dataset in the UK

Different subsites- head and neck and breast will be
pilot sites
In collaboration with NCIN and Public Health England
Summary




2 studies using routine data validated against manually
collected data demonstrating potential of analysing
national databases for clinically relevant outcomes
Can be automated and less resource-intensive than audit
Algorithms can be tailored for other cancer subsites
(GBM study under way)
Third phase study
Questions?