We could know the results before the trial starts

1
We could know the results before the trial starts
JG Chase Centre for Bio-Engineering University of
Canterbury New Zealand
T Desaive Cardiovascular Research
Center University of Liege Belgium
2
The problem (1)

• Critically ill patients can be defined by the
  high variability in response to care and
  treatment.
• Variability in outcome arises from
• variability in care
• variability in the patient-specific response to
  care.
• The greater the variability, the more difficult
  the patients management and the more likely a
  lesser outcome becomes.

3
The problem (2)

• Recent increase in importance of protocolized
  care to minimize the iatrogenic component due to
  variability in care.
• BUT protocols are potentially most applicable
  to groups with well-known clinical pathways and
  limited comorbidities, where a one size fits
  all approach can be effective.
• Those outside this group may receive lesser
  care and outcomes compared with the greater
  number receiving benefit.
• Need to try to reduce the component due to
  inter- and intra-patient variability in response
  to treatment.
• Model-based methods to provide patient-specific
  care

4
A Well Known Story

• Application Tight glycaemic control (TGC)
• TGC can improve outcomes BUT difficult to achieve
  without hypoglycemia
• In-silico simulated clinical trials (Virtual
  trials) can increase safety and save time cost
• Enable the rapid testing of new TGC intervention
  protocols and analysing control protocol
  performance
• Used to simulate a TGC protocol using a virtual
  patient profile identified from clinical data and
  different insulin and nutrition inputs.
• Virtual trials can help predicting outcomes of
  both individual intervention and overall trial
  cohort

5
The Model

• Physiologically Relevant Model

6
Model

• Model-based SI
• Whole-body insulin sensitivity
• Overall metabolic balance, including net effect
  of
• Altered endogenous glucose production
• Peripheral and hepatic insulin mediated glucose
  uptake
• Endogenous insulin secretion
• Has been used to guide model-based TGC in several
  studies
• Provides a means to analyse the evolution and
  hour-to-hour variability of SI in critically ill
  patients
• Enables prediction of variability in future

7
Virtual Trials

• Virtual Trials

8
Self Cross Validation

• The Glucontrol study randomised patients to two
  arms
• Group A Treated with Protocol A (intensive
  insulin protocol)
• Group B Treated with Protocol B (conventional
  insulin protocol)
• Two clinically matched cohorts that received
  different insulin treatments.
• Test the assumption of independence of clinical
  inputs (insulin) and patient state (insulin
  sensitivity parameter SI)

9
Virtual Trials Repeat Whole Trial Results

• CDFs of BG for clinical Glucontrol data and
  virtual trials on a (whole cohort)
• Validates the idea that virtual patients can
  INDEPENDENTLY capture effects of different
  treatment (cross validation results)

Excellent correlation and thus, the Virtual
patients are very good for tight control where
Insulin and safety risks are higher

• Very good match. Small 0.1-0.2 mmol/L shift due
  to several factors
• B patients often receive zero insulin
• Model assumptions on endog insulin
• Model assumptions on EGP
• Protocol non-compliance clinically
• Model assumptions have no effect on A case where
  exogenous inputs are higher and impact is thus
  less

10
Virtual Trials Per-Patient Results

• Median Difference Per-Patient (Self Validation)
• Variation due to model and compliance errors
  95 less than 15 error

Median BG is within 10 for 85-95 of patients
11
Virtual Trials Predicted Outcome SPRINT

• SPRINT was simulated first in to show efficacy
• Clinical virtual results are almost identical
• Other protocols were simulated for comparison
• Shows ability to know the answer first or at
  least have a lot of confidence

Virtual trials of 160 patients vs first 160
clinical patients (20k hours)
12
Virtual Trials Predicted Outcome STAR

• Virtual Trials on 371 virtual patients from
  SPRINT data but given STAR model-based protocol
• Clinical virtual results are almost identical
  for first 2000 hours
• Virtual trials done before clinical data for
  first 15 patients shown here
• Improvements using STAR and models is evident
  compared to SPRINT
• Shows ability to optimise with confidence in
  silico (safely and first)

13
Summary

• Virtual patients are effective and accurate
  portrayals of outcome, regardless of input used
  to make them.
• For a whole cohort
• For a specific patient
• Virtual patients and in-silico virtual trial
  methods are validated with cross validation with
  independent Glucontrol data
• Overall, we have a highly effective and
  physiologically representative model for design,
  analysis and real-time application of TGC
  protocols, in silico before they are implemented
  clinically!
• Methods readily extensible to other drug delivery
  problems to help predicting trials outcomes.

14
Conclusion

• Model-based methods can be used to develop safely
  and quickly BEFORE trials so
• We know the outcome ahead of time

15
Acknowledgments
The Hungarians Dr Balazs Benyo, Dr Levente
Kovacs, Mr Peter Szalay and Mr Tamas Ferenci, Dr
Attila Ilyes, Dr Noemi Szabo, ...