The present invention relates to methods of determining a risk of mortality of a cancer patient. The methods find application in contexts where quantification of such risk is beneficial, such as in assessment of anti-cancer therapy and/or in selection of patients for patient management regimes, treatment plans, or clinical trials.
Factors influencing life expectancy are important in public health. In oncology, prediction of patient survival is instrumental for optimal patient management. By understanding which variables are prognostic, we gain insights into disease biology, and are able to improve design, conduct, and data analysis of clinical trials and real-world data. Current research on prognostic factors in oncology is largely based on comparatively small sample sizes and studies one risk factor at a time. Examples of such research include the following:
Existing prognostic scores in common use are constructed from a relatively small number of risk factors. Such scores include the following:
The recent UK biobank initiative (Sudlow, C. et al. UK Biobank: An Open Access Resource for Identifying the Causes of a Wide Range of Complex Diseases of Middle and Old Age. PLoS Med. 12, e1001779 (2015)) constitutes an important addition to data availability. The initiative was used by Ganna and Ingelsson (Ganna, A. & Ingelsson, E. 5-year mortality predictors in 498 103 UK Biobank participants: a prospective population-based study. The Lancet 386, 533-540 (2015)) to investigate life expectancy in a population-based sample of ~500,000 participants and construct a mortality risk score outperforming the Charlson comorbidity index (Charlson, M. E., Pompei, P., Ales, K. L. & C.Ronald, M. A new method of classifying prognostic comorbidity in longitudinal studies: Development and validation. J Chronic Dis. 373-383 (1987) doi:10.1016/0021-9681(87)90171-8).
It is an object of the invention to improve quantitative determination of mortality risk.
According to an aspect, there is provided a computer-implemented method of determining a risk of mortality of a cancer patient, the method comprising: receiving patient data representing information about a cancer patient; using a mathematical model of mortality risk to determine a risk of mortality of the cancer patient based on the received patient data; and outputting the determined risk of mortality, wherein: the mathematical model models a relationship determined from training data between the risk of mortality and values of at least the following model parameters:
Thus, a method is provided in which a risk of mortality is determined and output. The output may be provided in the form of a numerical score. The numerical score may be output as data or shown on a display. The determination is performed using a model based on a specific selection of parameters that provides an optimized balance between high sensitivity and specificity on the one hand and manageable computational and data collecting demands on the other. As data included herein demonstrates, the core choice of 16 model parameters was found to provide sensitivity and specificity significantly above what has been observed to be possible in alternative prior art prognostic scores. Improved performance is feasible by including still further model parameters but the improvement that is possible is relatively small in comparison with the advantage already obtained relative to the alternative prior art prognostic scores. The selection of the core 16 model parameters is thus demonstrated to provide the good balance between sensitivity and specificity on the one hand and manageable computational and data collecting demands on the other.
The model was developed by the inventors performing prognostic modelling of overall survival (OS) on cohort data from the Flatiron Health database (122,694 patients). The inventors validated their results in two independent clinical studies. They examined demographic, clinical, hematology and blood chemistry parameters (focusing on routinely collected data), cancer diagnosis, and alongside real-world mortality as endpoint and assessed survival time from the first line of treatment.
The model parameters represent selected factors reflecting both tumor biology and patient characteristics and the determined score is shown to have a prognostic value that strongly outperforms contemporary risk scores. The score can be used to improve risk stratification, patient matching, and interpretation of clinical study results for early- and late-stage oncology drug development.
The inventors found that the determined score correlated not only with OS but also with particular early study dropout. Hence, using the determined score to exclude very high-risk patients may help protect patients from unnecessary exposure to study procedural burden and potential adverse events.
The inventors also found a surprisingly clear increase in the determined score towards death. During dose escalation in first-in-human studies it is a plausible hypothesis that the determined score will increase under inefficient drug dosages or treatments, but will stay stable or even improve under effective treatment. In this way, group level analysis of the time course of the determined score by treatment arm or drug dosage can give valuable insights on potential drug efficacy. Increase of determined score might, in combination with death events, be used as a surrogate end-point in clinical studies.In an embodiment, the mathematical model models the relationship determined from training data between the risk of mortality and values of the model parameters (i)-(xvi) and at least the following parameters:
In an embodiment, the mathematical model models the relationship determined from training data between the risk of mortality and values of the model parameters (i)-(xxvi) and at least the following parameter: (xxvii) alanine aminotransferase enzymatic activity level in serum or plasma.
This particular selection of 27 model parameter has been found by the inventors to provide a particularly beneficial balance of performance. The inventors observe significantly improved separation of Kaplan-Meier survival curves in comparison with RMHS.
In an embodiment, the mathematical model models the relationship determined from training data between the risk of mortality and values of the model parameters (i)-(xxvii) and at least the following parameters:
The inventors have found that adding these two extra model parameters provides improvement in sensitivity and/or specificity without excessive increases in computational and/or data collecting burden.
The inventors have demonstrated that the approach is robust to missingness in data at levels up to about 10 missing values of parameters. Missing values may be replaced, for example, by mean values of the parameters of interest, taken from the training data for the model. In an embodiment, the patient data comprises values for each of at least 16 of the model parameters where at least model parameters (i)-(xxvi) are used. Where only model parameters (i)-(xvi) are used, the approach will typically be robust to missingness in data at levels up to about 3 or 4 missing values of parameters.
In an embodiment, the model comprises a weighted sum of deviations in the patient data from mean values of the model parameters in the training data, and the model is formed by determining the weightings from the training data. Forming the model in this way is computationally efficient and provides high performance. The trained model can be represented in a compact form and applied with modest computational resources.
Embodiments of the invention will now be described, by way of example only, with reference to the accompanying drawings.
The following abbreviations are used herein:
In the following, methods are described for determining a risk of mortality of a cancer patient. The determined risk may be represented as a score (i.e. a quantitative measure, such as a real number). This score may be referred to herein as a Roche Prognostic Score (RoPro). The model that is used to obtain the score may be referred to herein as a RoPro model.
Methods of the present disclosure may be computer-implemented. Each step of the disclosed methods may be performed by a computer in the most general sense of the term, meaning any device capable of performing the data processing steps of the method, including dedicated digital circuits. The computer may comprise various combinations of computer hardware, including for example CPUs, RAM, SSDs, motherboards, network connections, firmware, software, and/or other elements known in the art that allow the computer hardware to perform the required computing operations. The required computing operations may be defined by one or more computer programs. The one or more computer programs may be provided in the form of media or data carriers, optionally non-transitory media, storing computer readable instructions. When the computer readable instructions are read by the computer, the computer performs the required method steps. The computer may consist of a self-contained unit, such as a general-purpose desktop computer, laptop, tablet, mobile telephone, smart device (e.g. smart TV), etc. Alternatively, the computer may consist of a distributed computing system having plural different computers connected to each other via a network such as the internet or an intranet.
In some embodiments, the mathematical model models a relationship determined from training data between the risk of mortality and values of a selected group of model parameters. The model is thus a trained model. The training data may comprise historical patient data from a database. Any of the various known ways of training a mathematical model using training data may be used.
The inventors used training data obtained from the Flatiron Health database (flatiron: https://flatiron.com). They performed a retrospective cohort analysis using electronic health records (EHRs) from the Flatiron Health database. The Flatiron Health database contains demographically and geographically diverse longitudinal data from over 280 oncology clinics in the USA. Institutional review board approval of the study protocol was obtained before study conduct and included a waiver of informed consent. Data from the February 2020 data release were extracted, including patient demographics and clinical data (e.g., type of cancer, disease stage, comorbidities, medication prescriptions, routine blood biomarkers). Patient records with missing first line of treatment information and patient records/variables with high missing data rates were excluded. Missing values in the final analysis data set were imputed.
In some embodiments, the model is formed by performing multivariable Cox regression analysis on the training data for a plurality of subjects, preferably at least 1000 subjects.
The inventors used a Cox proportional hazard model (Cox DR. Regression Models and Life-Tables. Journal of the Royal Statistical Society. Series B (Methodological) Vol. 34). Risk of mortality was investigated as overall survival (OS). Survival time was calculated from the start of the first line of treatment (defined as T0) to the event of “death” (death from all causes), as coded in the real-world mortality table (Stock, C., Mons, U. & Brenner, H. Projection of cancer incidence rates and case numbers until 2030: A probabilistic approach applied to German cancer registry data (1999-2013). Cancer Epidemiol 57, 110-119 (2018)) in the Flatiron Health database. For cohorts comprising only patients with advanced/metastatic disease, the first line of systemic treatment for advanced/metastatic disease is classed as first line in the database. Censored follow-up times were computed as days elapsed from T0 to the date of the patient’s last documented clinic contact. The patient’s last available pre-T0 measurement was used for modelling. A time limit of 30 days prior to T0 was applied. See Connolly, J. G., Schneeweiss, S., Glynn, R. J. & Gagne, J. J. Quantifying bias reduction with fixed-duration versus all-available covariate assessment periods. Pharmacoepidemiol. Drug Saf. 28, 665-670 (2019).
The inventors discovered from early experimentation with a smaller data version from 2018 that there was very little evidence that the routine variables available in Flatiron Health are of varying importance across cohorts. The inventors therefore built a pan-indication score (representing a determined risk of mortality).
For model comparison, the following were considered for censored data and are referred to below:
All analyses were conducted with the statistical analysis package R (R Core Team (2013). R: A language and environment for statistical computing. R Foundation for Statistical Computing, Vienna, Austria. URL http://www.R-project.org/).
Model selection was performed via a Family-wise error rate controlled (FWER-controlled) backward selection procedure on the Cox regression model, starting from 46 initial variables. The variable with least impact on model performance was iteratively removed until all remaining parameters yielded a measurable improvement in concordance index and generalized-r2 and were significant at α1=0.05/46=0.0011 (Bonferroni-correction). By construction, the procedure controls the family-wise error rate at α=0.05, i.e., all parameters are significant after adjustment for multiple testing. In parallel, the inventors derived a regularized Lasso model with 10-fold cross-validation (Simon N, Friedman J, Hastie T, Tibshirani R. Regularization Paths for Cox’s Proportional Hazards Model via Coordinate Descent. J. Stat. Softw. Artic. 2011; 39(5):1-13). The model performance is visualized in
For continuous variables, the Cox model yields hazard ratios (HRs) per unit of the investigated variable (model parameter); e.g., the HR for age is per age difference of one year. For better comparability across variables, the “4SD-HR”, the HR of a patient with a particularly high variable value (equal to mean + 2 standard deviations (SD)) versus one with a low value (mean - 2SD), is provided as descriptive measure.
Overall, 122,694 patients across 17 cancer-type-specific cohorts were eligible for analysis. Table 1 summarizes the patient characteristics.
1Indicates if log-transformation was performed, “2” indicates log2-transformation. All measures are given after the transformation.
2Hazard ratio on scale of variable (= per measurement unit, if applicable after log-transformation), together with 95% confidence interval
3Hazard ratio of patient with parameter value of mean+2SD compared to patient with value of mean-2SD
44SD-HR of lasso model
5Mean value in Flatiron Health data (after transformation, if applicable)
6Standard deviation in Flatiron Health data (after transformation, if applicable)
7Coded 1=male, 0=female
8Coded 1=History of smoking, 0= No history or unknown
9Not included in lasso model
10Neutrophils-lymphocytes-ratio
11TNM coding
12Included only in lasso model
After quality control, data were available for 46 general measurements, overall missing rate prior to imputation 21%, and 99 further cohort-specific measures (cytogenetics, biomarkers). Median survival time was 19.0 months (95% confidence interval [CI] 18.8-19.2).
Initial unadjusted analysis identified 44 variables significantly associated with OS. The backward selection model selected 27, while the cross-validated Lasso model selected 28 independently contributing variables, of which 26 coincided (Table 1). In both cases, plots of the concordance index by the number of included variables showed no evidence for an earlier saturation of the models with fewer variables (
The 27 variables found to be predictive may be used as the model parameters for the model of mortality risk. In the specific example described here, the resulting determined risk is output as a prognostic score referred to as a RoPro score. The score is based on a weighted sum of the patients’ differences from the respective reference means over the variables (see Table 1 for details). An example of the RoPro score may be determined from the following: 0.00948(age-67.15) + 0.14162(sex-0.502) + 0.19521(smoking-0.576) + 0.02833(Number of metastatic sites-0.897) - 0.04178(Hgb-12.087) + 0.19097(urea nitrogen-2.777) - 6e-04(platelets-267.423) + 1.0619(calcium-2.231) + 0.06098(glucose-4.733) -0.11658(lymphocytes-leukocytes-ratio-2.953) + 0.19019(ALP-4.582) - 0.00896(protein-68.888) - 0.05113(ALT-3.008) - 0.03988(albumin-37.851) + 0.08189(bilirubin-(-0.773)) -0.02462(chloride-101.434) + 0.10671(monocytes-(-0.548)) - 0.0543(eosinophils-leukocytes-ratio-0.307) + 1.22733(LDH-1.694) + 0.42372(heart rate-4.402) - 0.39878(SBP-(4.846)) - 0.02083(oxygen-96.487) + 0.20066(ECOG-0.84) + 0.0905(NLR-1.156) -0.17076(BMI-3.301) + 0.13122(AST-ALT-ratio-0.092) + 0.081(TumorStage-3.098).
The 27 model parameters are as follows:
In the example above, the model of mortality risk models a relationship determined from training data between the risk of mortality and values of all of the model parameters (i)-(xxvii), which is found to provide particularly high performance. However, acceptable performance may also be obtaining using a smaller number of the available model parameters.
In some embodiments, the model models the relationship determined from the training data between the risk of mortality and the values of at least 16, optionally at least 18, optionally at least 20, optionally at least 22, optionally at least 24, optionally all of, the model parameters (i)-(xxvi). In some embodiments, the model models the relationship determined from the training data between the risk of mortality and the values of at least 16, optionally at least 18, optionally at least 20, optionally at least 22, optionally at least 24, of the model parameters (i)-(xxvii).
In some embodiment, the following additional model parameters are used: (xxviii) level of eosinophils in blood; and (xxix) diastolic blood pressure.
The use of these extra model parameters increases the model complexity and data requirements, but may provide further performance improvements. In embodiments of this this type, the model may model the relationship determined from the training data between the risk of mortality and the values of at least 16, optionally at least 18, optionally at least 20, optionally at least 22, optionally at least 24, optionally all of, the model parameters (i)-(xxix).
For a particular patient, it may be that values for some of the model parameters used by the model are not available. The inventors have found that the model can still provide a reliable score in these circumstances up to a missingness of about 10 values. Thus, the patient data should preferably comprise values for each of at least 16, optionally at least 18, optionally at least 20, optionally at least 22, optionally at least 24, of the model parameters used by the model. Alternatively, the patient data comprises values for all of the model parameters (i)-(xxvi). Alternatively, the patient data comprises values for all of the model parameters (i)-(xxvii). Alternatively, the patient data comprises values for all of the model parameters (i)-(xxix).
As mentioned above, in some embodiments the mathematical model models the relationship between the risk of mortality and values of at least 16 model parameters. In one class of embodiment, the 16 model parameters include the following:
In the case where only these 16 model parameters are used, the model may be referred to as the 16 variable reduced model or “ROPRO16” model.
In some embodiments, the model is expanded to 26 parameters by including the following parameters:
In some embodiments, the model is expanded to 27 parameters by including the following parameter: (xxvii) alanine aminotransferase enzymatic activity level in serum or plasma.
In some embodiments, the model is expanded to 29 parameters by including the following parameters:
As mentioned above, in some embodiments the determination of the mortality risk comprises calculating a numerical score representing the mortality risk. In some embodiments, the determination of the mortality risk further comprises comparing the calculated score to one or more predetermined threshold values, or to calculated scores for other cancer patients. For example, it may be determined that the mortality risk is “high” if the calculated score is above a predetermined threshold value, or above an average score calculated for a selected group of cancer patients. The mortality risk may be determined as “low” if the calculated score is below a predetermined threshold value, or below an average score calculated for a selected group of cancer patients.
The score thus provides a practical quantitative tool that represents a physical state of a patient. The score is obtained the same way for different patients or for the same patient at different times. The score can thus be used as an objective, quantitative basis for comparing different patients between each other and/or for comparing states of the same patient at different times. The score thus provides a quantitative tool which can be used in various ways to support objective decision making in relation to providing health care for patients and/or to facilitate extraction of objective information from studies that involve the patients.
In some embodiments, the model comprises a weighted sum of deviations in the patient data from mean values of the model parameters in the training data, and the model is formed by determining the weightings from the training data.
In an embodiment, the model is formed by: assigning a respective weighting, wi, to each of the model parameters, and determining a respective mean, mi, of values of each model parameter over the training data. The determination of the mortality risk may then comprise calculating a numerical score according to the following formula:
where wi is the weighting of the i-th model parameter, mi is the mean of the i-th model parameter, and mij is the value of the i-th model parameter for a j-th cancer patient for whom the score is to be calculated.
In an embodiment, each of the weightings wi comprises a natural logarithm of the Hazard Ratio (HR), such that the score is given as follows:
where HR (xi) is the HR estimated for variable i, mi is the variable mean in the training data (e.g. Flatiron Health database), and mij is the value of patient j at variable i, i ∈ I.
The determined risk of mortality (e.g. as a score) may be used in various situations as mentioned above.
In one class of embodiment, a method of assessing an anti-cancer therapy that uses the determined score may be provided. The method comprises determining a risk of mortality of a patient at plural different times while the patient is receiving the anti-cancer therapy, by performing the method of determining a mortality risk according any of the embodiments of the present disclosure at each of the plural times, and analysing the resulting determined risks to determine an efficacy of the anti-cancer therapy.
Alternatively or additionally, a method of selecting cancer patients for treatment with an anti-cancer therapy that uses the determined score may be provided. The method comprises determining a risk of mortality of a candidate patient using the method of determining a mortality risk according to any of the embodiments of the present disclosure. The determining risk is then used to decide whether to select each candidate patient.
Further examples of applications of the determined risk (e.g. score) are discussed below.
The performance of an example implementation of the method of determining a risk of mortality is discussed in this section. The example implementation uses the 27 model parameters (i)-(xxvii). In this case, the determined risk is represented as a score that is referred to as a RoPro score or simply “RoPro”.
Individual patient RoPro scores (derived by inputting their measurements for each of the variables into the formula) ranged from -3.49 to 4.12, 99% in [-2.21; 2.09]. Higher scores indicate higher risk.
The model in this example was found to strongly outperform the RMHS with respect to all model performance indicators (r2=0.32, C-index=0.747, 3-month-AUC=0.82 vs. r2=0.03, C-index=0.54, 3-month-AUC=0.58).
The RoPro score showed strong performance in all 17 cohorts, based on highly consistent variable (model parameter) estimate sizes and directions across cohorts. With cohort-specific re-estimated variable weights, ln(HR(xi), strong improvement of model performance was seen for chronic lymphocytic leukemia (CLL) (r2=0.13, C-index=0.704, 3-month-AUC=0.82 for general RoPro in CLL; r2=0.17, C-index=0.74, 3-month-AUC=0.84 for CLL-specific RoPro).
For validation purposes, the (RoPro) model was applied retrospectively, e.g. without re-estimation of the variable weights, to patients from two independent clinical studies. Here, the RoPro score, as a single variable, was fitted using Cox regression. The first analysis was based on a phase I first-in-human study, BP29428 (NCT02323191), which investigated the combination of emactuzumab, a monoclonal antibody targeting the receptor for colony stimulating factor-1 (CSF1) (Gomez-Roca, C. A. et al. Phase I study of emactuzumab single agent or in combination with paclitaxel in patients with advanced/metastatic solid tumors reveals depletion of immunosuppressive M2-like macrophages. Ann. Oncol. 30, 1381-1392 (2019)) and atezolizumab, a monoclonal antibody targeting PDL1, in patients (n=216) with locally advanced or metastatic solid tumors not amenable to standard treatment. The second analysis used data from the phase III OAK study, which evaluated efficacy and safety of atezolizumab monotherapy versus single agent docetaxel in participants (n=1187) with locally advanced or metastatic non-small-cell lung cancer (NSCLC) after failure of platinum-containing chemotherapy (Rittmeyer, A. et al. Atezolizumab versus docetaxel in patients with previously treated non-small-cell lung cancer (OAK): a phase 3, open-label, multicentre randomised controlled trial. Lancet 389, 255-265 (2017)).
First, the inventors could replicate the correlation of RoPro with OS in the phase I study BP29428 (n=216, P=6.08×10-15, r2=0.23, C-index=0.81). Patients with RoPro>1.18 (cut-off equal to 90%-quantile in Flatiron Health, n=8) had particular poor OS prognosis (HR 11.29; 95% CI 5.52-23.12). Three-month-survival ROC-AUC values were comparable to that in Flatiron Health (
Second, the correlation of the RoPro with OS survival was also replicated in the phase III study OAK (n=1,187, P=8.38×10-60, r2=0.20, C-index=0.69). Patients with RoPro>0.82 (cut-off equal to 90%-quantile in Flatiron Health for dedicated advanced NSCLC RoPro, n=34) had poorer OS prognosis (HR 3.65; 95% CI 2.59-5.12). Area under the curve values again outperformed RMHS (
In the phase I study, the inventors saw a correlation of high RoPro (>1.18, see definition above, n=9) with study dropout. All high RoPro patients left the study very early, either due to death (n=5) or progressive disease (n=4). Only one patient could receive more than one treatment cycle, although all patients had an ECOG performance status value of 1 at baseline (a study inclusion criterion). Typically, study protocols require a life expectancy of >12 weeks for enrollment since drug efficacy is potentially masked in patients with a particularly bad prognosis. However, the RoPro analysis demonstrates that with current practice it is not always possible to identify all high-risk patients a priori. With the RoPro available, however, respective information can retrospectively be included to exclude patients from analysis or to perform stratified analysis.
In the phase III study, the inventors assessed potential impact of using the prognostic score as a covariate parameter in comparison of OS between the treatment arms. In plain analysis, an HR of 0.794 (95% CI 0.690-0.913, P=0.0012) was observed for atezolizumab versus docetaxel, in accordance with published results (Connolly, J. G., Schneeweiss, S., Glynn, R. J. & Gagne, J. J. Quantifying bias reduction with fixed-duration versus all-available covariate assessment periods. Pharmacoepidemiol. Drug Saf. 28, 665-670 (2019)). Although not required in a randomized study, the inventors adjusted the efficacy analysis for baseline RoPro. The inventors could confirm the efficacy gain with atezolizumab and obtained a HR of 0.760 (95% CI 0.666-0.875, P=0.00013) in the RoPro adjusted analysis.
Finally, the inventors made an initial assessment whether changes of RoPro over time might be indicative of later events. An ad hoc visualization is given in
In this analysis of 122,694 patients from 17 cancer cohorts, 27 independent OS risk factors were identified and used to define a prognostic model, the RoPro. It showed a previously unseen level of correlation with OS. Validation analyses using data from two independent clinical studies confirmed these results. To the inventors’ knowledge, the RoPro is based on the largest dataset used for prognostic modelling to date, overcoming typical limitations. Furthermore, the inventors have proven within a classic statistical framework that a multitude of prognostic factors have an independent contribution to risk modelling, after adjustment for correlation between the risk factors. The results were highly consistent when using different modeling approaches (backward-selection and regularized regression).
The RoPro may comprise 27 clinical parameters combined into a single score and can be readily applied to new data sets. A substantial improvement in performance over alternative scores from models such as the RMHS, the IPI or the IMDC has been demonstrated.
By construction, all RoPro variables are typically available in routine clinical practice in the U.S. The majority of variables are also available in clinical practice in other countries and clinical studies conducted by pharmaceutical industry. Patient-specific missingness of five to ten variables is tolerated for computing a patient’s RoPro.
Finally, the RoPro can be applied across cancer indications other than the 17 used to build the model, as demonstrated by the performance in BP29428, where 40% of patients had cancer indications not available in Flatiron Health.
The inventors’ analysis in study BP29428 (emactuzumab/atezolizumab combination, NCT02323191) showed that RoPro correlated not only with OS but also with particular early study dropout. Hence, using RoPro to exclude very high-risk patients may help protect patients from unnecessary exposure to study procedural burden and potential adverse events, although this needs to be balanced against the benefits of allowing such patients access to potentially effective novel drugs.
Landmark analysis in the OAK study demonstrated that an early increase/decrease of the RoPro is indicative of later survival events. While biologically plausible, we note that this observation is not self-evident. A possible alternative scenario would have been one of early-stage and late-stage prognostic factors, where early-stage factors do not deteriorate from a certain time point onwards and late-stage mechanisms are indicative of developments very close to death event.
The increase of the RoPro towards death is a feature that generates possible applications of high potential. First, during dose escalation in first-in-human studies it is plausible hypothesis that the RoPro will increase under inefficient drug dosages or treatments, but will stay stable or even improve under effective treatment. In this way, group level analysis of RoPro time course by treatment arm or drug dosage can give valuable insights on potential drug efficacy. Increase of RoPro might, in combination with death events, be used as a surrogate end-point in clinical studies.
Continuous monitoring of the RoPro over time could accompany treatment decision making. The inventors’ analysis of OAK study data indicated that baseline high-risk patients whose scores in addition worsen over time are unlikely to benefit from therapy, the majority of such patients died within very short time. Therefore, patients with increasing RoPro could be switched to another treatment. Observing stable scores or even slightly improving scores might indicate treatment benefits and could be used, along with many other considerations, in decisions about continuing treatment.
The fact that RoPro works well on “new” cancer indications as exemplified by its application to the Phase I study, which recruited many patients from outside the US, increases confidence that the RoPro is robust across geographic regions and ethnicities.
While uncertainties and inaccuracies are expected to be encountered with retrospective real-world data analysis, the inventors’ findings suggest that the power gain obtained by sample size overcomes potential bias.
The results make comparisons with prior art alternative prognostic scores, including RMHS, IPI, IMDC, and ECOG (all mentioned above).
Arrangements of the disclosure are defined in the following numbered clauses.
1. A method of determining a risk of mortality of a cancer patient, the method comprising:
2. The method of clause 1, wherein the patient data comprises values for each of at least 16 of the model parameters used by the model.
3. The method of clause 1 or 2, wherein the patient data comprises values for all of the model parameters (i)-(xxvi).
4. The method of any clauses 1-3, wherein the model parameters further include: (xxvii) alanine aminotransferase enzymatic activity level in serum or plasma.
5. The method of clause 4, wherein the patient data comprises values for all of the model parameters (i)-(xxvii).
6. The method of clause 4 or 5, wherein the model parameters further include:
7. The method of clause 6, wherein the patient data comprises values for all of the model parameters (i)-(xxix).
8. The method of any of clauses 1-7, wherein the mathematical model models a relationship determined from the training data between the risk of mortality and values of all of the model parameters (i)-(xxvi).
9. The method of any of clauses 4-7, wherein the mathematical model models a relationship determined from the training data between the risk of mortality and values of all of the model parameters (i)-(xxvii).
10. The method of clause 6 or 7, wherein the mathematical model models a relationship determined from the training data between the risk of mortality and values of all of the model parameters (i)-(xxix).
11. The method of any of clauses 1-10, wherein the determination of the mortality risk comprises calculating a numerical score representing the mortality risk.
12. The method of clause 11, wherein the determination of the mortality risk further comprises comparing the calculated score to one or more predetermined threshold values, or to calculated scores for other cancer patients.
13. The method of any of clauses 1-12, wherein the model comprises a weighted sum of deviations in the patient data from mean values of the model parameters in the training data, and the model is formed by determining the weightings from the training data.
14. The method of any of clauses 1-13, wherein the model is formed by performing multivariable Cox regression analysis on the training data for a plurality of subjects, preferably at least 1000 subjects.
15. The method of any of clauses 1-14, wherein:
16. A method of assessing an anti-cancer therapy, comprising determining a risk of mortality of a patient at plural different times while the patient is receiving the anti-cancer therapy, by performing the method of any of clauses 1-15 at each of the plural times, and analysing the resulting determined risks to determine an efficacy of the anti-cancer therapy.
17. A method of selecting cancer patients for treatment with an anti-cancer therapy, comprising determining a risk of mortality of a candidate patient using the method of any of clauses 1-15 and using the determining risk to decide whether to select each candidate patient.
18. The method of any of clauses 1-17, wherein the method is carried out by a computer.
19. A computer program comprising instructions which, when the program is executed by a computer, cause the computer to carry out the method of any of clauses 18.
Number | Date | Country | Kind |
---|---|---|---|
20177361.1 | May 2020 | EP | regional |
Filing Document | Filing Date | Country | Kind |
---|---|---|---|
PCT/EP2021/064062 | 5/26/2021 | WO |