Bayesian forecasting utilizing bleeding information to support dose individualization of factor VIII

João A. Abrantes (1), Alexander Solms (2), Dirk Garmann (3), Elisabet I. Nielsen (1), Siv Jönsson (1), Mats O. Karlsson (1)

(1) Department of Pharmaceutical Biosciences, Uppsala University, Uppsala, Sweden, (2) Bayer, Berlin, Germany, (3) Bayer, Wuppertal, Germany

Objectives:
Model-based PK-guided dose individualization of factor VIII (FVIII) replacement therapy has been increasingly encouraged [1,2]. Yet, mounting evidence shows large phenotypic differences in bleeding between patients due to multiple components besides plasma FVIII activity [3,4].

The aim of this work was to employ a pharmacokinetic-repeated time-to-event (PK-RTTE) model to contrast different sources of patient information in their ability to predict future occurrence of bleeds in severe haemophilia A patients receiving prophylactic FVIII replacement therapy.

Methods:
Data and model
Dose, covariate, observed plasma FVIII activity and bleeding time data collected over 6 to 12 months during the Long-Term Efficacy Open-Label Program in Severe Hemophilia A Disease (LEOPOLD) clinical trials were used in this evaluation [5-7].

A previously developed integrated population PK-RTTE-FREM model for FVIII was used for Bayesian forecasting [3].

Bayesian forecasting and bleeding probabilistic forecast
Empirical Bayes estimates (EBEs) of PK and hazard parameters were estimated based on the information observed from the start of the LEOPOLD study up to the end of each consecutive 24-hour period, i.e. repeatedly for each patient up to the end of study. The estimation was performed using three information scenarios:

1. PK - plasma FVIII activity observations;
2. Bleed - time of bleed, or lack of bleed, during each day;
3. All - plasma FVIII activity observations, time of bleeds and covariate information.

Subsequently, the longitudinally estimated individual bleeding hazard was used to derive the individual forecasted probability of having a bleed in the upcoming 24-hour period (P_i(bleeding)).

The effect of the duration of the Bayesian observation period was also assessed by estimation based on the past 15 days, 1, 2, 3 or 6 months, to investigate the trade-off between longer periods and the most up-to-date information.

Bleeding predictive performance assessment
The predictive performance of the different information scenarios was assessed by comparing P_i(bleeding) with the individual observed occurrence of a bleed on the forecasted day, using separation plots, receiver operating characteristic (ROC) and precision-recall analyses [8-10]. The optimal threshold of P_i(bleeding) in the ROC analyses was determined by the Youden’s J statistics.

Results:
In total, 101 bleeds were observed in 51 patients aged <12 years, and 530 bleeds in 121 patients aged ≥12 years, and days with observed bleeds were ~1% of the forecasted days.

For the group <12 years, the expected number of bleeds over the study period was 66 (PK), 96 (Bleed), and 90 (All), and for ≥12 years it was 218 (PK), 461 (Bleed) and 500 (All). Separation plots showed a sharper increase in P_i(bleeding) associated to days when bleeds occurred for Bleed and All compared to PK, for both age groups. The ROC curves showed that Bleed had a predictive power comparable to All, and both were superior to PK (Table 1).

Table 1 - Summary statistics of the ROC analyses for the different information scenarios in patients ≥12 years.

The differences between scenarios in the group <12 years followed the same trends, with an AUC of 0.67 (0.61-0.72) for PK, 0.74 (0.69-0.79) for Bleed and 0.77 (0.73-0.81) for All. The results of ROC analyses were confirmed by the precision-recall analyses, with PK closer to the performance of a random classifier.

Using Bleed, patients with a high bleeding risk required shorter observation periods to inform the EBEs, namely, between 60 and 90 days prior to the EBEs estimation. No advantage was found to use only the most up-to-date information to estimate the EBEs.

Conclusions:
A PK-RTTE-FREM model-based forecasting approach considering the efficacy endpoint of interest (bleeds) under prophylactic treatment has been developed. Using observed data to contrast sources of information to be used in Bayesian forecasting, this work suggests that individual bleed information adds value to the optimization of prophylactic therapy in severe haemophilia A. Further steps to optimize the proposed tool for FVIII dose adaptation in the clinic are required.

References:
[1] Berntorp E. If you know you will also see: population pharmacokinetics is the way to personalize and optimize prophylaxis in hemophilia. J Thromb Haemost. 2017;15(6):1103-5.
[2] Iorio A, Blanchette V, Blatny J, et al. Estimating and interpreting the pharmacokinetic profiles of individual patients with hemophilia A or B using a population pharmacokinetic approach: communication from the SSC of the ISTH. J Thromb Haemost. 2017;15(12):2461-5.
[3] Abrantes JA, Solms A, Garmann D, Nielsen EI, Jonsson S, Karlsson MO. Integrated modelling of factor VIII activity kinetics, occurrence of bleeds and individual characteristics in haemophilia A patients using a full random effects modelling (FREM) approach. PAGE 27. 2018; abstract 8646 [www.page-meeting.org/?abstract=8646].
[4] Iorio A, Iserman E, Blanchette V, et al. Target plasma factor levels for personalized treatment in haemophilia: a Delphi consensus statement. Haemophilia. 2017;23(3):e170-e9.
[5] Saxena K, Lalezari S, Oldenburg J, et al. Efficacy and safety of BAY 81-8973, a full-length recombinant factor VIII: results from the LEOPOLD I trial. Haemophilia. 2016;22(5):706-12.
[6] Kavakli K, Yang R, Rusen L, et al. Prophylaxis vs. on-demand treatment with BAY 81-8973, a full-length plasma protein-free recombinant factor VIII product: results from a randomized trial (LEOPOLD II). J Thromb Haemost. 2015;13(3):360-9.
[7] Ljung R, Kenet G, Mancuso ME, et al. BAY 81-8973 safety and efficacy for prophylaxis and treatment of bleeds in previously treated children with severe haemophilia A: results of the LEOPOLD Kids Trial. Haemophilia. 2016;22(3):354-60.
[8] Greenhill B, Ward MD, Sacks A. The Separation Plot: A New Visual Method for Evaluating the Fit of Binary Models. American Journal of Political Science. 2011;55(4):991-1002.
[9] Fawcett T. An introduction to ROC analysis. Pattern recognition letters. 2006;27(8):861-74.
[10] Davis J, Goadrich M. The relationship between Precision-Recall and ROC curves. Proceedings of the 23rd international conference on Machine learning. 2006:223-240.

Reference: PAGE 28 (2019) Abstr 8944 [www.page-meeting.org/?abstract=8944]

Oral: Clinical Applications

PDF poster/presentation (click to open)