How to predict ADAS-Cog scores
The method uses machine learning on tensor-encoded clinical trial data to predict ADAS-Cog scores for Alzheimer's patients under unknown treatments, reducing bias and enabling efficient clinical trials with personalized treatment designs.
Patent Information
- Application Number
- JP2024537017
- Authority / Receiving Office
- JP · JP
- Patent Type
- Applications
- Current Assignee / Owner
- Priority Date
- 2022-08-03
- Filing Date
- 2022-12-13
- Publication Date
- 2024-12-05
- Estimated Expiration
- Not applicable · inactive patent
AI Technical Summary
Current methods for predicting cognitive decline in Alzheimer's disease, as measured by the ADAS-Cog scale, are inefficient and prone to bias due to dropout patients in clinical trials, and lack personalized treatment designs.
A method using machine learning to generate synthetic models from tensor-encoded clinical trial data, allowing prediction of ADAS-Cog scores for patients under unknown treatment regimens by leveraging data from similar patients, and enabling data-efficient clinical trial designs.
Reduces bias in clinical trials by predicting ADAS-Cog scores for dropout patients and allows for more efficient clinical trial designs with fewer subjects, while providing personalized treatment recommendations.
Smart Images

Figure 2024545277000001_ABST
Abstract
Description
[Technical field]
[0001] FIELD OF THEINVENTION The present invention relates to a method for predicting the level of cognitive function, and in particular, but not exclusively, to a method for predicting the cognitive score of a patient or patient cohort on the Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog). [Background technology]
[0002] background Alzheimer's disease is a neurological disorder in which excess proteins accumulate in the brain, impairing neuronal function and ultimately causing cell death. The disease is characterized by a continuous progression, although the rate of progression varies from person to person.
[0003] The Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog) was developed in the 1980s to track the rate of progression of Alzheimer's disease in individuals, and in particular to determine and track the level of cognitive and non-cognitive functioning of individuals. Since then, it has also been used as an aid in the diagnosis of Alzheimer's disease, for example in pre-dementia populations. The ADAS-Cog is now considered the gold standard for evaluating the effectiveness of treatments for dementia, including Alzheimer's disease.
[0004] Diagnosis of Alzheimer's disease, and subsequent monitoring of a subject's mental performance under specific treatment, is an active and multifaceted area of research and is often the subject of clinical trials.
[0005] Pharmaceutical companies are increasingly using advanced analytics to help conduct clinical trials.
[0006] This trend is supported by the growing interest in personalized medicine and the rapid development of modern statistical techniques and computational resources with the advent of machine learning (Shah et al., 2019;Bhatt, 2021). Improving the efficiency of clinical trials by adapting their design and analysis is of particular interest. Moreover, dropouts (i.e., patients who do not undergo the full course of a clinical trial) are a major source of bias in trials and can be a major reason for trial failure (Little et. al., 2012).
[0007] The present invention has been conceived in light of the above considerations. Summary of the Invention [Means for solving the problem]
[0008] Summary of the Invention In a first aspect, there is provided a method for predicting a cognitive score of a patient or patient cohort on the Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog), comprising: obtaining data relating to a plurality of patients or patient cohorts, the data including information regarding longitudinal trajectories of ADAS-Cog scores for the plurality of patients or patient cohorts over time, each patient or patient cohort receiving a treatment regimen selected from a plurality of treatment regimens; encoding the data into tensors across patients or patient cohorts, time, and treatment plans; generating a synthetic model of the target patient or target patient cohort using machine learning processes and tensors; Using the synthetic model to predict the ADAS-Cog score of a target patient or a target patient cohort under a target treatment plan selected from a plurality of treatment plans; A method is provided that includes:
[0009] The inventors have found that this data on ADAS-Cog trajectories of multiple patients when undergoing multiple different treatment regimens (e.g., from clinical trials for the treatment of Alzheimer's disease) surprisingly shows a significant latent linear structure of the induced tensors, even though Alzheimer's disease patients have heterogeneous health status trajectories. This is particularly surprising at the individual patient level. This means that, since there is a limited number of normalized patient (or patient cohort) profiles, every patient can be represented as a linear combination of these profiles in terms of their ADAS-Cog scores. As a result, this allows us to predict the ADAS-Cog trajectory associated with any patient under the same treatment regimen by leveraging information from other patients who have undergone a particular treatment regimen.
[0010] In this way, the ADAS-Cog score, or the longitudinal trajectory of the ADAS-Cog score, of a target patient or target patient cohort under an unknown treatment regimen (i.e., a treatment regimen not received by the target patient) can be estimated.
[0011] It should be understood that a targeted treatment is generally an unknown treatment regimen, i.e., a treatment regimen that has not been received by the target patient, but for which it is desired to predict an outcome if the patient were to receive the targeted treatment regimen.
[0012] Next, optional features are described, which can be applied alone or in combination with any aspect of the present disclosure.
[0013] It should be understood that a synthetic model is a model that uses available data to create a synthetic trajectory (of ADAS-Cog scores over time for a patient / patient cohort and treatment combination) that is not seen in the available data.
[0014] Preferably, the data on a plurality of patients or patient cohorts is data from a clinical trial for treating Alzheimer's disease.
[0015] It should be understood that the method may not be limited to predicting ADAS-Cog scores, but may also be applied to predicting other scores for predicting neurocognitive status or decline, such as the Mini-Mental State Examination (MMSE), Global Deterioration Scale (GDS), Montreal Cognitive Assessment (MoCA), etc.
[0016] Therefore, more generally, the method may be a method for predicting the neurocognitive status of a patient or patient cohort. The data obtained may therefore include information on the longitudinal trajectory of the neurocognitive status of a plurality of patients or patient cohorts over time. The target patient or target patient cohort may be undergoing an actual treatment plan, which is different from the target treatment plan. In particular, the target patient / cohort may be undergoing an "actual" treatment plan, and the outcome of that first treatment plan may be observed. The outcome of the target patient / cohort under a second, different "target" treatment plan that the target patient / cohort has not undergone may be predicted using the synthetic model. Here, the term "outcome" is understood as the ADAS-Cog score of the target patient / cohort at time t, or the ADAS-Cog trajectory of the target patient / cohort over time.
[0017] The tensor may be a third-order tensor.
[0018] The data in the tensor may include, for each patient of multiple patients / cohorts, data regarding (i) the ADAS-Cog trajectory of each patient / cohort over time (e.g., a baseline measurement of each patient's ADAS-Cog score (e.g., at t=0) and the patient's / cohort's ADAS-Cog score at multiple subsequent time points / visits); and (ii) the treatment regimen that each patient / cohort received.
[0019] One or more of the multiple treatment plans may include one or more different neuropharmacological interventions, such as, for example, medicines or therapies for the treatment of Alzheimer's disease. One treatment plan of the multiple treatment plans may correspond to no treatment (e.g., patients receiving this treatment plan are included in the control group). The different treatment plans may include different administration regimes of treatment / medication. For example, a first treatment plan may correspond to a high dose of treatment combined with the use of standard treatment, a second treatment plan may correspond to a high dose of treatment combined with the absence of standard treatment, a third treatment plan may correspond to a low dose of treatment combined with the use of standard treatment, and a fourth treatment plan may correspond to a low dose of treatment combined with the absence of standard treatment.
[0020] A synthetic model of a target patient or target patient cohort can be generated using data relating to the target patient or target patient cohort, which may include data regarding (i) the ADAS-Cog trajectory of the target patient / cohort over time (e.g., a baseline measurement of the target patient / cohort's ADAS-Cog score (e.g., at t=0) and the target patient / cohort's ADAS-Cog score at multiple subsequent time points / visits) and (ii) the actual treatment regimen received by each patient.
[0021] Where the data relates to a patient cohort rather than an individual patient, each ADAS-Cog score may be the average (eg mean) ADAS-Cog score across the cohort.
[0022] Preferably, the synthetic model may be generated using a machine learning process based at least in part on a set of observed data in a tensor that corresponds to the same treatment plan as the actual treatment plan received by the target patient or target patient cohort. In particular, the synthetic model may be learned at least in part from a set of observed data from a tensor that relates to the actual treatment plan received by the target patient / cohort. This data may include a set of observed pairs (time, treatment plan) of data that the target patient / cohort and the patient data in the tensor have in common. In other words, the machine learning process may be based on a set of (study visit, treatment plan) data pairs for the target patient / cohort and patients in the tensor under the same (actual) treatment plan, corresponding to the same visit time (e.g., day).
[0023] The machine learning process may involve determining the minimum norm linear relationship between data relating to a target patient / cohort and a set of observed data in a tensor that is relevant to the actual treatment plan.
[0024] The machine learning process may include performing a Principal Component Regression (PCR) analysis. In particular, the machine learning process may include performing a PCR between data related to a target patient / cohort and a set of observed data in a tensor that corresponds to an actual treatment plan received by the target patient / cohort. The PCR process may define a unique minimum-norm linear relationship between the data related to the target patient / cohort and the set of observed data in a tensor that relates to the actual treatment plan.
[0025] Alternatively, other known machine learning algorithms (parametric or non-parametric) can be used in the machine learning process to learn the relationship between data related to the target patient / cohort and the set of observed data in tensors associated with the actual treatment plan, for example, neural networks or random forests can be used.
[0026] Optionally, the method may include validating the synthetic model for the target patient / cohort. This may involve determining a training error of the learned synthetic model and then proceeding to predict the ADAS-Cog score of the target patient / cohort using the synthetic model only if the determined training error meets a predefined criterion / threshold. In this way, it can be verified that the synthetic model has been trained to a satisfactory standard. Furthermore, this step also verifies that the underlying data is likely to meet the desired properties to provide an accurate estimate.
[0027] Predicting the ADAS-Cog score for the target patient / cohort under the target treatment plan may be based on the trained synthetic model and the set of data in the tensor corresponding to the target treatment plan.
[0028] The prediction may be a prediction of the ADAS-Cog trajectory of a target patient / cohort under a target treatment regimen over time, or a prediction of the ADAS-Cog score at a particular time point t.
[0029] Optionally, the method may include using the synthetic model to predict the ADAS-Cog score of the target patient / cohort under each treatment regimen of the multiple treatment regimens.
[0030] As described above, the data may be from a clinical trial to treat Alzheimer's disease, e.g., to evaluate the efficacy of a neuropharmacological intervention (such as a pharmaceutical drug) in treating Alzheimer's disease. The plurality of patients may be a treatment group that has been diagnosed with Alzheimer's disease. The plurality of patients may include patients that have been diagnosed with Alzheimer's disease.
[0031] The above method may be computer-implemented. For example, the method may be implemented on one or more computing devices. The one or more computing devices may be, for example, one or more computers, servers, cloud-based devices.
[0032] The method may further include using the predicted ADAS-Cog score of the target patient or target patient cohort under the target treatment regimen in the outcome of a clinical trial.
[0033] As a first example, the method may be for predicting missing data in a clinical trial due to dropouts (patients who do not complete the clinical trial). Thus, the target patient may be a dropout from the clinical trial, and the method may further include using the predicted ADAS-Cog score of the target patient as an outcome for the target patient in the clinical trial results.
[0034] In this way, even if a patient drops out of the study and cannot observe the entire ADAS-Cog trajectory over time, and therefore the outcome of the study, the patient's ADAS-Cog score over time can be predicted by utilizing the observations of other patients in the study. Dropouts are generally a large source of bias in studies and can be a major reason for study failure. This method reduces this bias.
[0035] The method of this aspect may include a combination of some, all, or none of the preferred and optional features described above.
[0036] As another example, the methods can be used to design data-efficient clinical trials for personalized treatment at the patient cohort level.
[0037] In particular, in a second aspect of the invention there is provided a method for processing data for a clinical trial to treat a neurocognitive disease using the method described above, comprising the steps of: obtaining data relating to at least two patient cohorts, the first patient cohort including a first subset of patients receiving a first treatment regimen and a second subset of patients receiving a second treatment regimen, the second patient cohort including the first subset of patients receiving the first treatment regimen and a second subset of patients receiving a third treatment regimen, the data including information regarding longitudinal trajectories of ADAS-Cog scores of the patient cohorts over time; encoding the data into tensors across patient cohorts, time, and treatment plans; generating a first synthetic model of the first patient cohort and a second synthetic model of the second patient cohort using a machine learning process and tensors; Using the first synthetic model to predict ADAS-Cog scores for a first patient cohort under a third treatment regimen, and using the second synthetic model to predict ADAS-Cog scores for a second patient cohort under a second treatment regimen. A method is provided that includes:
[0038] In this way, the first / second composite model can predict the ADAS-Cog score (or ADAS-Cog trajectory over time) for any patient cohort under unknown treatment. This allows for the design of clinical trials that include many fewer patients / subjects, since the subject outcomes can be predicted by leveraging the corresponding outcomes from a smaller subset of subjects undergoing clinical trials. This results in better data-efficient clinical trials.
[0039] The clinical trial may be, for example, a trial to treat Alzheimer's disease or behavioral frontotemporal dementia (FTD).
[0040] The tensor may be a third-order tensor.
[0041] The data in the tensor may include, for each patient in the patient cohort, data regarding (i) each patient's ADAS-Cog trajectory over time (e.g., each patient's baseline measurement of their ADAS-Cog score (e.g., at t=0), and the patient's ADAS-Cog score at multiple subsequent time points / visits); and (ii) the treatment regimen that each patient received.
[0042] The treatment plans may include one or more different neuropharmacological interventions, such as, for example, medicines or therapies for the treatment of Alzheimer's disease. One of the treatment plans may correspond to no treatment (e.g., patients receiving this treatment plan are included in the control group). The different treatment plans may include different administration regimes of the treatment / medication. For example, a first treatment plan may correspond to a high dose of treatment combined with the use of standard treatment, a second treatment plan may correspond to a high dose of treatment combined with the absence of standard treatment, and a third treatment plan may correspond to a low dose of treatment combined with the use of standard treatment. A fourth treatment plan may correspond to a low dose of treatment combined with the absence of standard treatment.
[0043] The first / second synthetic models may be generated using data relating to the first / second patient cohorts, respectively, which may include data regarding (i) the ADAS-Cog trajectories of the first / second patient cohorts over time (e.g., baseline measurements (e.g., at t=0) of the ADAS-Cog scores of the first / second patient cohorts and the ADAS-Cog scores of the first / second patient cohorts at multiple subsequent time points / visits) and (ii) the actual treatment regimens received by patients within the respective patient cohorts.
[0044] Each ADAS-Cog score may be the average (eg, mean) ADAS-Cog score across the cohort.
[0045] Preferably, the first / second synthetic models may be generated using a machine learning process based at least in part on a set of observed data in tensors that correspond to the same treatment plan as the actual treatment plan received by the patients in the first / second patient cohorts, respectively. For example, the first synthetic model may be generated using a machine learning process based at least in part on a set of observed data in tensors that correspond to the first treatment plan (because a subset of patients in both the first and second patient cohorts receive the first treatment plan).
[0046] The machine learning process may include determining a minimum norm linear relationship between data relating to the first / second patient cohort and a set of observed data in a tensor associated with a subgroup of patients within the first / second cohort that received the first treatment plan.
[0047] The machine learning process may include performing a Principal Component Regression (PCR) analysis. In particular, the machine learning process may include performing PCR between data related to the first / second patient cohort and a set of observed data in a tensor corresponding to a first treatment regimen received by a subset of patients in the first / second patient cohort, respectively. The PCR process may define a unique minimum norm linear relationship between the data related to the first / second patient cohort and a set of observed data in a tensor related to a first treatment regimen received by a subset of patients in the first / second patient cohort, respectively.
[0048] Alternatively, other known machine learning algorithms (parametric or non-parametric) can be used in the machine learning process to learn the linear relationships, for example, neural networks or random forests.
[0049] Optionally, the method may include validating the first and / or second synthetic models. This may include determining a training error of the learned first / second synthetic models, and then proceeding to predict the ADAS-Cog score of the target patient / cohort using the synthetic model only if the determined training error meets a predefined criterion / threshold. In this way, it can be verified that the synthetic model has been trained to a satisfactory standard. Moreover, this step also verifies that the underlying data is likely to meet the desired properties to provide an accurate estimate.
[0050] The step of predicting the ADAS-Cog scores of the first / second patient cohorts under the third / second treatment plans may be based on the trained first / second composite models and the sets of data in tensors corresponding to the third / second treatment plans, respectively.
[0051] The prediction may be a prediction of the ADAS-Cog trajectory of the 1st / 2nd patient cohort under the 3rd / 2nd treatment regimen over time, or a prediction of the ADAS-Cog score at a particular time point t.
[0052] The above method may be computer-implemented. For example, the method may be implemented on one or more computing devices. The one or more computing devices may be, for example, one or more computers, servers, cloud-based devices.
[0053] The method of this aspect may include a combination of some, all, or none of the preferred and optional features described above.
[0054] In a third aspect, there is provided a system including one or more processors and a memory, the memory including machine-executable instructions that, when executed on the one or more processors, cause the one or more processors to perform the method of the first aspect.
[0055] The memory may include machine-executable instructions which, when executed on the processor, cause the processor to perform the method of the first aspect including any one of the optional features described with reference to the first aspect or, where compatible, any combination thereof.
[0056] Accordingly, there is provided a system for predicting a cognitive score for a patient or a cohort of patients on the Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog), comprising: a data acquisition means configured to acquire data relating to a plurality of patients or patient cohorts, the data comprising information regarding longitudinal trajectories of ADAS-Cog scores for the plurality of patients or patient cohorts over time, each patient or patient cohort receiving a treatment regimen selected from a plurality of treatment regimens; a tensor generator configured to encode the data into tensors across patients or patient cohorts, time, and treatment regimens; a synthetic model generator configured to generate a synthetic model of a patient or patient cohort using machine learning processes and tensors; a predicting means configured to predict an ADAS-Cog score for a patient or a cohort of patients under a target treatment regimen of a plurality of treatment regimens using the synthetic model; A system is provided comprising:
[0057] In some embodiments, the system may comprise, for example, one or more computers, servers, or cloud-based devices.
[0058] The system can be a system for processing data for a clinical trial to treat Alzheimer's disease, the data acquisition means is configured to collect data from at least two patient cohorts, the first patient cohort including a first subset of patients receiving a first treatment regimen and a second subset of patients receiving a second treatment regimen, the second patient cohort including the first subset of patients receiving the first treatment regimen and a second subset of patients receiving a third treatment regimen, and the data includes information regarding longitudinal trajectories of ADAS-Cog scores of the patient cohorts over time; a tensor generator configured to encode the data into tensors across patient cohorts, time, and treatment plans; a synthetic model generator configured to generate a first synthetic model for a first cohort of patients and a second synthetic model for a second cohort of patients using the machine learning process and the tensors; The predicting means is configured to predict the ADAS-Cog score for the first patient cohort under a third treatment regimen using the first composite model, and to predict the ADAS-Cog score for the second patient cohort under the second treatment regimen using the second composite model.
[0059] Thus, the system may be configured to carry out the method of the second aspect including any one of the optional features described with reference to the second aspect or any combination thereof where compatible.
[0060] According to a fourth aspect, there is provided a non-transitory computer readable storage medium comprising machine executable instructions which, when executed on a processor, cause the processor to perform a method of the first or second aspect including any one or, where compatible, any combination of the optional features described with reference to the first or second aspect.
[0061] According to a fifth aspect there is provided a computer program comprising executable code which, when executed on a computer, causes the computer to perform the method of the first or second aspect.
[0062] According to a sixth aspect there is provided a computer readable storage medium storing a computer program comprising code which, when executed on a computer, causes the computer to perform the method of the first or second aspect.
[0063] The present invention includes combinations of the described aspects and preferred features unless such combinations are expressly not permitted or explicitly avoided.
[0064] Drawing Overview BRIEF DESCRIPTION OF THE DRAWINGS Embodiments and experiments illustrating the principles of the present invention are discussed below with reference to the accompanying drawings. [Brief description of the drawings]
[0065] [Figure 1] FIG. 1 includes two graphs showing the spectrum of third-order tensors of clinical trial data at (i) the patient cohort level under mode 2 deployment and (ii) the individual patient level under “high-use”, respectively. [Diagram 2] FIG. 1 includes two tables showing actual test data and data-efficient test data, respectively. [Diagram 3] FIG. 13 is a graph showing the error associated with data-efficient testing. [Figure 4] FIG. 1 includes plots showing the mean absolute error and standard error of imputing missing data for clinical trial dropouts across different treatments and different numbers of observed treatments using a method according to an aspect of the present disclosure, compared to other baseline measures. [Diagram 5] FIG. 1 includes plots showing a comparison of observed densities associated with clinical trial adherers and estimated densities associated with dropouts predicted using a method according to an aspect of the present disclosure. [Figure 6] 1 is a flow diagram of a method for predicting the cognitive score of a patient or patient cohort on the Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog). [Figure 7] FIG. 7 is a schematic diagram of a system for implementing the method of FIG. 6. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS
[0066] Detailed Description of the Invention Aspects and embodiments of the present invention are discussed below with reference to the accompanying drawings. Further aspects and embodiments will be apparent to those skilled in the art. All documents mentioned herein are incorporated by reference.
[0067] We performed an analysis of the above method on real clinical trial data and found that it can be used with synthetic intervention (SI) estimators designed to recover low-rank tensors with non-random (i.e., causally validated) missing data.
[0068] Clinical Trial Overview Data were obtained from Phase III clinical trials for Alzheimer's disease (TRx-237-005 and TRx-237-015). 1,162 patients participated in these trials. Baseline measurements were collected for each patient before the study began. This included the patient's sex, age, baseline ADAS-Cog, and Mini-Mental State Examination (MMSE) scores.
[0069] Below, X i is expressed as a vector of baseline measurements associated with patient i.
[0070] Across two Phase III trials, patients were randomized to treatment with 8, 150, 200, or 250 mg / day of LMTM (developed by TauRx Therapeutics), but a lack of dose response was observed in both trials. Further population pharmacokinetic analysis showed a steep concentration-response relationship for steady-state plasma levels in the 8 mg / day range. Patients were then classified into those with "low" and "high" plasma levels using a threshold based on the lower limit of the amount assayed on day 1. This analysis showed a significant difference in cognitive decline and brain atrophy in the "high" plasma level group (Schelter et. al., 2019). Furthermore, each patient participated in the study regardless of whether they were already using standard treatment. Thus, D=4 possible combinations were obtained; namely, "low-use", "low-nonuse", "high-use", and "high-nonuse". The study spanned five visits, with 13 weeks between each visit. At each visit t, the progressive ADAS-Cog score for every patient i
number
[0071] Encoding clinical trial data as tensors Clinical trial data, both at the patient cohort level and at the individual patient level, can be encoded as third-order tensors.
[0072] To demonstrate this, we first encode the clinical trial data at the patient cohort level into a third-order tensor. Before the start of the trial, patients are assigned a baseline measurement X i The study was clustered into six cohorts based on the observed study visit outcomes (i.e., Y tiFor each patient cohort, we observed the ADAS-Cog trajectories under every treatment combination (except for one cluster in which no patients received the “low-absence” treatment). We then modeled the data as a third-order tensor
number
[0073] Next, we analyzed data at the individual patient level, rather than at the cohort level. We analyzed the available individual patient ADAS-Cog trajectory data as a third-order tensor
number
number
[0074] The spectral profile of Y1 was analyzed by examining its mode 1 and mode 2 evolution, as shown in plot 10 of Figure 1. Further details on the tensor evolution can be found in section 7 of Agarwal, A et al., (2020). As shown in plot 10, Y1 was found to have a substantially low canonical polynomial (CP) tensor rank, i.e., 1. Y2 was also analyzed. Plot 12 of Figure 1 shows the spectral profile of Y2 under the “High-Use” treatment. dThe spectral profile of Y2 is shown in Fig. 1. d It was also shown to be almost rank 1.
[0075] Y1 and Y2 d Since both , and are low-rank tensors, it turns out that there is a significant underlying linear structure in the induced tensor / matrix. In particular, this means that since there is a limited number of normal patient or patient cohort profiles, all patients can be represented as a linear combination of these profiles in terms of their ADAS-Cog scores. This is surprising given the heterogeneous health status trajectories of Alzheimer's patients. This therefore turns out that the present invention allows us to predict the ADAS-Cog trajectory associated with any patient under a particular treatment by leveraging information from other patients who have undergone that treatment using a known synthetic intervention estimator.
[0076] Trajectory prediction with synthetic intervention Without loss of generality, the aim of the above method is to predict the ADAS-Cog score for patient (or cohort) i at visit t under treatment d, which we denote by Y ti (d) This is expressed as:
[0077] A two-step synthetic intervention estimator is used to estimate the heterogeneous ADAS-Cog trajectories for each patient under each treatment. An example of such a two-step SI estimator is disclosed in Agarwal, A et al., (2020).
[0078] The first step is to compare patients i and J d A machine learning model trained on common observations with patients in
number
[0079] The notation used to aid understanding is set out below.
[0080] Ω d ⊂[T]×[D] is the probability distribution between patients i and J d The set of observed (study visit, treatment) pairs that patients i and J have in common, i.e., patients i and J with the same visit days and the same treatment. d Let T denote the set of (study visit, treatment) pairs for patients and d =|Ω d |Let's say.
[0081]
number
number
number
[0082] Similarly, J. d For patients in
number
number
number
[0083] The singular value decomposition is
number
number
[0084] The two-step SI estimator should be retained.
number
[0085] As mentioned above, the first step of the SI estimator is to use a machine learning model
number
number
[0086] The second step of the SI estimator is J d The goal is to simulate the ADAS-Cog score of patient i under treatment d at time t by leveraging the observed outcomes of patients in and the learned synthetic model.
number
[0087] Data-Efficient Clinical Trials for Personalized Treatment As stated, the methods described above can be used to design data-efficient clinical trials for personalized treatment at the patient cohort level.
[0088] To demonstrate this, we defined the "low-use" drug combination as the control, since patients in this cohort essentially received the current standard of care. To simulate a data-efficient study, the SI estimator described above allowed us to observe ADAS-Cog scores under control for all six patient cohorts. However, the SI estimator only allowed us to observe each patient cohort under one of the other three possible treatments (high-use, low-nonuse, high-nonuse). That is, we purposely presented data under these three treatments as a test set to see if the SI estimator could reproduce these trajectories. A graphical representation of the observed patterns for the actual study data can be seen in Table 14 of Figure 2, and a graphical representation of the observed patterns for the simulated data-efficient study data can be seen in Table 16 of Figure 2. In Figure 2, check marks in cells represent observed outcomes, and blank cells represent unobserved (i.e., presented) outcomes.
[0089] Given the sparsity of the patterns in Table 16 of Figure 2, the objective was to estimate the mean ADAS-Cog trajectory for every patient cohort under each of the two unobserved treatments. That is, for an unobserved treatment d, the aim is to estimate
number
[0090] To quantify the predictive accuracy of the SI estimator for any cohort i and treatment d, the randomized controlled trial (RCT) estimator was used as a baseline. The RCT estimator takes the average ADAS-Cog outcome across all patients who received treatment d as its predicted value. Thus, for example, to predict the ADAS-Cog outcome of cohort 3 in Figure 2 under high-use, the RCT estimator takes the average ADAS-Cog outcome of all patients in cohorts 1 and 2. When patients within a study are homogeneous, the RCT estimator becomes a powerful predictor. More formally, for any cohort i and treatment d, the squared error is defined as follows:
number
number
number
[0091] In the above setting, the SI estimator was applied to predict the ADAS-Cog trajectory for every patient cohort under unknown treatment. To mitigate randomness, the experiment was repeated 10 times with different treatment assignments for each iteration. The results can be seen in graph 18 of Figure 3.
[0092] Graph 18 in Figure 3 shows the squared error across each cohort under any treatment.
number
[0093] Across almost all cohorts i and treatments d, the median
number
[0094] This confirms that there is indeed large heterogeneity between the cohorts, and remarkably, the SI estimator accurately predicts outcome despite this heterogeneity and despite having access to only a small subset of the data.
[0095] Imputing missing values for dropouts in clinical trials Dropouts (i.e., patients who do not complete the entire clinical trial) are a major source of bias in trials and can be a major reason for trial failure (Little et. al., 2012). We found that the above method for predicting ADAS-Cog scores can be used to impute missing data values due to dropouts.
[0096] To validate the SI estimator described above, we simulated dropouts on observed patient data. To make the simulation realistic, the probability of a patient being a dropout was determined according to the change in the patient's ADAS-Cog score. This closely matches the reality that patients who tend not to respond to treatment are more likely to drop out.
[0097] The simulation was set up as follows:
[0098] T0 < T was selected as the number of hospital visits when no patient had dropped out. Based on the average change in the ADAS-Cog score of the patients from the baseline after the hospital visit T0, the patients were clustered into poor, moderate, and good responders. Poor responders are those patients with an average change from the baseline greater than one standard deviation above the mean across all patients. Good responders are those patients one standard deviation below the mean. Moderate responders are the remaining patients, i.e., those within one standard deviation. As described above, an increase in the ADAS-Cog score means the disease is worsening.
[0099] The probability of dropout after the hospital visit T0 was selected as 60%, 30%, and 10% for poor, moderate, and good responders, respectively. Since the number of patients was small under the low - non - use combination, the focus was on the other three combinations.
[0100] The aim was to estimate the average change from the baseline for each treatment across all patients (simulated compliers and dropouts).
Number
[0101] Accuracy was measured by the relative mean absolute error
Number
Number
Number
[0102] We compared the estimators to the baseline estimator used by Little et al., 2012. (i) Perfect case: using only data from adherents, μ (d) (ii) RCTs; (iii) LOCF (Last Observation Carried Forward): imputes missing data for dropouts using the last available outcome (commonly used in longitudinal studies); (iv) multivariate singular spectral analysis (mSSA): imputes missing data for dropouts using a spectral-based time series algorithm (this estimator has been demonstrated to outperform many state-of-the-art time series algorithms on standard benchmarks) (Agarwal, et. al., 2021).
[0103] T0 was varied from 1 to 4. This procedure was repeated 10 times for each T0, and the resulting mean absolute error and standard error are shown in plot 20 shown in FIG. 4. The standard error is shown in plot 20 as shaded bands. It was found that the SI estimator not only achieves low relative error, but also consistently outperforms existing baselines.
[0104] The SI estimation method described above can also be used to quantify the bias present due to actual dropouts in the actual clinical trial conducted by TauRx described above, by constructing a counterfactual ADAS-Cog trajectory for every dropout patient if they had remained.
[0105] Figure 5 shows the distribution of mean change from baseline for adherers (actually observed) and dropouts (this is the counterfactual distribution constructed from the SI estimator).
[0106] Although the dropout data are actually missing in the real trial data, and so the estimated density cannot be verified, the qualitative conclusions were consistent with the feedback received from the dropout patients. In particular, plot 22a suggests a natural progression in which dropouts were likely to have dropped out of the trial due to a poor response to the low-use combination. In fact, these were the patients who said that the combination they were assigned made them feel worse. A similar situation is shown in plot 22b, although the difference between adherers and dropouts is less pronounced. In comparison, plot 22c suggests that dropouts and adherers responded similarly to the high-non-use combination. In fact, many of the dropouts for this particular combination stated that they dropped out of the trial due to external circumstances, such as moving to an inconveniently far away location for a new job.
[0107] Thus, we have shown that the methods described above can be used to accurately impute missing data values due to dropouts.
[0108] Methods for predicting ADAS-Cog scores 6 illustrates a flow diagram of a method for predicting the ADAS-Cog score of a target patient under an unknown treatment plan. The method can be performed on one or more computing devices.
[0109] At S101, data from a clinical trial for the treatment of Alzheimer's disease is obtained (e.g., from an internal or external storage system). The data includes information regarding the longitudinal trajectory of ADAS-Cog scores of multiple patients over time and the treatment plan each patient received (e.g., the data includes ADAS-Cog score information, time information, and treatment plan information).
[0110] In S103, this data is encoded into a third-order tensor across patient, time, and treatment plan.
[0111] In the first step of the synthetic intervention estimator, a synthetic model is generated at S105. The synthetic model is of a target patient and is generated using a machine learning process. As shown in S106, optionally, the synthetic model is trained based on a set of observational data in a tensor corresponding to an actual treatment plan received by the target patient and data related to the target patient. For example, the machine learning process may include determining a minimum norm linear relationship between the data related to the target patient and the set of observational data in a tensor related to the actual treatment plan received by the target patient. In particular, the PCR process may define a unique minimum norm linear relationship between the data related to the target patient and the set of observational data in a tensor related to the actual treatment plan.
[0112] S107 is an optional step. In S107, the synthetic model is verified to check its accuracy. The training error of the learned synthetic model is calculated based on the data on the target patient and the synthetic model to test whether the synthetic model can accurately predict the observation data related to the target patient. If the calculated error is less than a predefined threshold, the method proceeds to S109. Otherwise, the method ends.
[0113] In S109, the ADAS-Cog trajectory of the target patient under the unknown treatment plan is predicted using the synthetic model and the set of data in tensors corresponding to the target treatment plan as shown in S110.
[0114] In this way, it is possible to predict the target patient's ADAS-Cog trajectory under a particular treatment regimen that the target patient has not yet completed. This prediction can be used as the outcome of the treatment regimen by the target patient in the clinical trial results, such as when the target patient drops out of the clinical trial. This prediction can also be used to provide more data-efficient clinical trials (e.g., by predicting some of the outcomes, fewer subjects are needed for the clinical trial). This prediction can also be used to design more data-efficient clinical trials or to prescribe personalized therapy, for example by predicting which treatment regimen may be more efficient.
[0115] Based on the predictions, subpopulation-treatment pairs that yield promising outcomes can be further investigated, while pairs that yield unfavorable outcomes can be discarded. In this way, the above method can be used as a recommendation engine for strategically deciding which experiments to prioritize and as a tool towards the development of precision medicine.
[0116] Since one or more predictions can be used to predict the outcome of a clinical trial, the methods described above can also be used in early termination applications, for example to decide whether to continue or stop a clinical trial based on the predicted results.
[0117] Further applications of the above method are described below.
[0118] The first application concerns uncontrolled single-arm trials, which may be common in the field of rare and / or orphan diseases. These disease areas may have small patient populations and no established standard of care. These clinical trials may have insufficient statistical sample sizes and may raise ethical concerns regarding the administration of placebos to patients. To address these issues, external control data may be used. This external control data may be used as data for multiple patients or patient cohorts in the manner described above. In this way, a synthetic model may be used to construct a synthetic twin of every patient in the study under control using a weighted combination of data from the external control patients. The cohort of synthetic twins may then form a synthetic control group.
[0119] Further applications focus on demonstrating the safety and efficacy of approved treatments in the real world. In particular, to provide evidence of safety, comprehensive monitoring of newly approved treatments may be performed to identify potential hazards that may be harmful to patients. This may be referred to as adverse event reporting and may be used to determine whether adverse events are caused by the newly approved treatment. Generally, evaluation involves observational studies in which the incidence of adverse events is compared between a patient population under the treatment regimen and a natural control group. Using the methods described above, an artificial control group may be constructed. In particular, for every patient under the treatment, a synthetic twin under control may be constructed from external (e.g., real-world) control data. If the incidence of adverse events in the observed treatment group is statistically higher compared to the synthetic control group, it may be determined that the treatment contributes to the adverse events. To demonstrate efficacy, evidence may be provided that the newly released treatment is competitive against an established standard treatment regimen. Using the methods described above, a synthetic cohort under the standard treatment regimen may be constructed. In particular, for every patient under the new treatment regimen, the methods described above may generate a synthetic twin under the standard treatment regimen. To assess the comparative effectiveness of a new treatment compared with a standard treatment, the therapy outcomes of two treatment arms can be compared.
[0120] The above-described methods can also be used for label extension. For example, in the context of off-label patients (e.g., individuals who are under-represented or unrepresented in randomized controlled trials), the above-described methods can be used to construct a synthetic twin for every off-label patient under the (newly announced) treatment plan as a weighted combination of data from on-label patients. This can effectively simulate a synthetic trial of off-label patients under the (newly announced) treatment plan. This can provide researchers with evidence to support or oppose the extension of treatment to off-label subpopulations.
[0121] Below, we provide an example generalized algorithmic overview of the method of Figure 6 for constructing synthetic data for a given target "unit" i if it were to undergo intervention d, where we assume that data related to unit i is missing, where "unit" may represent a patient or a patient cohort, and "intervention" may represent a treatment regimen.
[0122] 1. Data Preparation a) Prepare a database for each unit containing data relating to all interventions received by that unit. b) Query the database for all observations related to target unit i. c) Query the database for all observations relating to the "donor" unit that received intervention d. d) filtering the data to identify a collection of observations related to common interventions between target unit i and the associated donor units that maximizes the minimum of (i) the number of satisfactory observations and (ii) the number of filtered donor units; e) Create a new table consisting of only data related to common interventions (from step (d)) associated with the target unit and the filtered donor units. Denote these two data structures as y1 and X1, respectively. f) Create a new table consisting of only intervention-related data associated with the filtered donor units under intervention d. Denote this data structure as X2.
[0123] 2. Data verification through diagnostic tests a) Create a new table that concatenates X1, X2 column by column, i.e. X=[X1,X2]. b) Perform a singular value decomposition of X and examine its spectral profile to see if X exhibits low-dimensional structure. If X does not exhibit low-dimensional structure, it may be necessary to preprocess X (e.g., apply an autoencoder to X to identify a new low-dimensional representation of X). c) Conduct a subspace containment hypothesis test on X1, X2 (e.g., as detailed in section 5 of Agarwal et al., 2020). If the hypothesis test passes, we have strong diagnostic evidence that we can accurately create synthetic data for unit i under intervention d.
[0124] Note that, as discussed above in relation to Figure 1, data from clinical trials for the treatment of Alzheimer's disease generally have low tensor rank, and therefore there is strong diagnostic evidence that synthetic data for unit i under intervention d can be accurately created.
[0125] 3. Generate synthetic data a) Select a predefined training error tolerance ε. b) Following the synthetic data generation procedure described above (further detailed in section 3 of Agarwal et. al., 2020). (i) Model training: Run Principal Component Regression (PCR) between y1 and X1
number
number
number
number
number
number
[0126] 4. Diagnosing the accuracy of synthetic data a) As an added layer of evaluation, a "cross-validation" test can be performed to investigate whether the above steps have been successful in creating an observational data set. Each donor unit can be repeatedly assigned to be the target unit, and the remaining donor units form a donor group for that particular intervention. That is, one can observe the observations of the temporal target unit under intervention d (i.e. one has access to "synthetic" data to replicate). Then,
number
number
[0127] Figure 7 is a schematic diagram of a system 200 for implementing the method of Figure 6. The system 200 may comprise, for example, one or more computing devices.
[0128] The system 200 comprises a data acquisition means 202 configured to acquire data relating to a plurality of patients or patient cohorts, the data including information regarding longitudinal trajectories of ADAS-Cog scores of a plurality of patients or patient cohorts over time, each patient or patient cohort receiving a treatment regimen selected from a plurality of treatment regimens. The data may be acquired from an external storage medium 210 via a wired or wireless connection.
[0129] The system 200 also comprises a tensor generator 204 configured to encode the data acquired by the data acquisition means 202 into tensors across patients or patient cohorts, time, and treatment regimens.
[0130] The synthetic model generator 206 is configured to generate a synthetic model of a patient or patient cohort using a machine learning process and the tensors generated by the tensor generator 204. The predictor 208 is configured to predict an ADAS-Cog score for a patient or patient cohort under a target treatment plan of the multiple treatment plans using the synthetic model generated by the synthetic model generator 208. This predicted ADAS-Cog score may be output from the system, for example, by a wired or wireless connection.
[0131] The systems and methods of the above embodiments, in addition to the structural components and user interactions described, can be implemented in a computer system (particularly computer hardware or computer software).
[0132] The term "computer system" includes hardware, software, and data storage devices for implementing the system or performing the method according to the above-described embodiments. For example, a computer system may comprise a central processing unit (CPU), input means, output means, and data storage. Preferably, the computer system has a monitor for providing a visual output display. The data storage may comprise RAM, disk drives, or other computer-readable media. A computer system may include multiple computing devices connected by a network and capable of communicating with each other via the network.
[0133] The methods of the above embodiments may be provided as a computer program, or as a computer program product or computer readable medium carrying a computer program configured to perform the above-mentioned methods when executed on a computer.
[0134] The term "computer-readable medium" includes, but is not limited to, any non-transitory medium that can be read and accessed directly by a computer or computer system, including, but not limited to, magnetic storage media, such as floppy disks, hard disk storage media, and magnetic tape, optical storage media, such as optical disks and CD-ROMs, electrical storage media, such as RAM, ROM, and memory, including flash memory, as well as hybrids and combinations of the above, magnetic / optical storage media, and the like.
[0135] The features disclosed in the foregoing description, or in the appended claims, or in the accompanying drawings, either presented in a particular form, or presented in terms of means for performing a disclosed function or, where appropriate, in a method or process for obtaining a disclosed result, can be utilized individually or in any combination of such features to realise the invention in various forms thereof.
[0136] Although the present invention has been described in relation to the exemplary embodiments set forth above, many equivalent modifications and variations will be apparent to those skilled in the art upon reading this disclosure. Accordingly, the exemplary embodiments of the present invention set forth above are considered to be illustrative and not limiting. Various changes can be made to the described embodiments without departing from the spirit and scope of the present invention.
[0137] For the avoidance of doubt, the theoretical explanations provided herein are provided for the purpose of enhancing the understanding of the reader, and the inventors do not wish to be bound by any of these theoretical explanations.
[0138] The section headings used herein are for organizational purposes only and are not to be construed as limiting the subject matter described.
[0139] Throughout this specification, including the appended claims, unless the context otherwise requires, the words "comprises" and "including," and variations such as "comprising," "comprising," and "including," will be understood to imply the inclusion of a specified integer, step, or group of integers or steps, and not the exclusion of any other integer, step, or group of integers or steps.
[0140] It should be noted that, as used in this specification and the appended claims, the singular forms "a," "an," and "the" include plural referents unless the context clearly dictates otherwise. Ranges may be expressed herein as from "about" one particular value and / or to "about" another particular value. When such a range is expressed, another embodiment includes from the one particular value and / or to the other particular value. Similarly, when values are expressed as approximations, by use of the antecedent "about," it will be understood that the particular value forms another embodiment. The term "about" in reference to numerical values is optional and may mean, for example, ±10%.
[0141] References Numerous publications are cited above in order to more fully describe and disclose the present invention and the prior art to which this invention pertains. Full citations for these references are set forth below. Agarwal, A., Shah, D., Shen, D.: Synthetic interventions (2020), arXiv:2006.07691 [econ.EM] Bhatt, A.: Artificial intelligence in managing clinical trial design and conduct: Man and machine still on the learning curve? Perspectives in Clinical Research 12(1), 1 (2021) Little, R.J., D’Agostino, R., Cohen, M.L., Dickersin, K., Emerson, S.S., Farrar, J.T., Frangakis, C., Hogan, J.W., Molenberghs, G., Murphy, S.A., Neaton, J.D., Rotnitzky, A., Scharfstein, D., Shih, W.J., Siegel, J.P., Stern, H.: The prevention and treatment of missing data in clinical trials. New England Journal of Medicine 367(14), 1355-1360 (2012). https: / / doi.org / 10.1056 / NEJMsr1203730, pMID: 23034025 Schelter BO, Shiells H, Baddeley TC, Rubino CM, Ganesan H, Hammel J, Vuksanovic V, Staff RT, Murray AD, Bracoud L, Riedel G, Gauthier S, Jia J, Bentham P, Kook K, Storey JMD, Harrington CR, Wischik CM. Concentration-Dependent Activity of Hydromethylthionine on Cognitive Decline and Brain Atrophy in Mild to Moderate Alzheimer’s Disease. J Alzheimers Dis. 2019;72(3):931-946. doi: 10.3233 / JAD-190772. PMID: 31658058; PMCID: PMC6918900. (2019) Shah, P., Kendall, F., Khozin, S., Goosen, R., Hu, J., Laramie, J., Ringel, M., Schork, N.: Artificial intelligence and machine learning in clinical development: a translational perspective. NPJ digital medicine 2(1), 1-5 (2019)
Claims
1. 1. A method for predicting a cognitive score for a patient or patient cohort on the Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog), comprising: obtaining data relating to a plurality of patients or patient cohorts, the data including information regarding longitudinal trajectories of ADAS-Cog scores for the plurality of patients or patient cohorts over time, each patient or patient cohort receiving a treatment regimen selected from a plurality of treatment regimens; encoding the data into tensors across patients or patient cohorts, time, and treatment regimens; generating a synthetic model of a target patient or target patient cohort using a machine learning process and said tensors; using the synthetic model to predict an ADAS-Cog score for a target patient or a target patient cohort under a target treatment regimen selected from the plurality of treatment regimens; The method includes:
2. The method of claim 1 , wherein the synthetic model of the target patient / cohort is generated using data relating to the target patient or target patient cohort.
3. The method of claim 1 or 2, wherein the tensor is a third-order tensor.
4. The method of any one of claims 1 to 3, wherein the target patient or target patient cohort is receiving an actual treatment regimen, and the actual treatment regimen is different from the target treatment regimen.
5. 5. The method of claim 4, wherein the synthetic model is generated using a machine learning process based at least in part on a set of observed data in the tensor that corresponds to the same treatment regimen as the actual treatment regimen received by the target patient / cohort.
6. 6. The method of claim 5, wherein the machine learning process comprises determining a minimum norm linear relationship between the data relating to the target patient / cohort and the set of observed data in the tensor associated with the actual treatment plan.
7. The method of any one of claims 1 to 6, wherein the machine learning process comprises performing a Principal Component Regression (PCR) analysis.
8. 8. The method of any one of claims 1 to 7, further comprising validating the synthetic model for the target patient or target patient cohort by determining a training error of the generated synthetic model, and then proceeding to predict the ADAS-Cog score for the target patient or target patient cohort using the synthetic model only if the determined training error meets a predefined criterion.
9. 9. The method of any one of claims 1 to 8, wherein predicting the ADAS-Cog score of the target patient / cohort under the target treatment plan is based on the synthetic model and a set of data in the tensor corresponding to the target treatment plan.
10. The method of any one of claims 1 to 9, wherein the data relating to the plurality or patient cohort is data from a clinical trial for treating Alzheimer's disease.
11. 11. The method of claim 10, further comprising using the predicted ADAS-Cog score of the target patient or target patient cohort under the target treatment regimen in the outcome of the clinical trial.
12. The method according to any one of claims 1 to 11, which is for predicting missing data in a clinical trial due to dropouts.
13. A method for processing data for clinical trials for treating a neurocognitive disorder using the method according to any one of claims 1 to 12, comprising: obtaining data relating to at least two patient cohorts, a first patient cohort including a first subset of patients receiving a first treatment regimen and a second subset of patients receiving a second treatment regimen, and a second patient cohort including the first subset of patients receiving said first treatment regimen and a second subset of patients receiving a third treatment regimen, said data including information regarding longitudinal trajectories of ADAS-Cog scores of said patient cohorts over time; encoding the data into tensors across patient cohorts, time, and treatment regimens; generating a first synthetic model of the first patient cohort and a second synthetic model of the second patient cohort using a machine learning process and the tensors; using the first synthetic model to predict ADAS-Cog scores for the first patient cohort under the third treatment regimen, and using the second synthetic model to predict ADAS-Cog scores for the second patient cohort under the second treatment regimen; The method includes:
14. the first and second composite models are generated using the machine learning process based at least in part on a set of observation data at the tensor corresponding to the first treatment plan. The method of claim 13.
15. predicting the ADAS-Cog score for the first patient cohort under the third treatment plan based on the first synthetic model and the set of data in the tensor corresponding to the third treatment plan; predicting the ADAS-Cog score for the second patient cohort under the second treatment plan based on the second synthetic model and the set of data in the tensor corresponding to the second treatment plan.
15. The method according to claim 13 or 14.
16. A computer-implemented method according to any one of claims 1 to 15.
17. 1. A system for predicting a cognitive score for a patient or patient cohort on the Alzheimer's Disease Assessment Scale-Cognitive Subscale (ADAS-Cog), comprising: a data acquisition means configured to acquire data relating to a plurality of patients or patient cohorts, said data comprising information regarding longitudinal trajectories of ADAS-Cog scores of said plurality of patients or patient cohorts over time, each patient or patient cohort receiving a treatment regimen selected from a plurality of treatment regimens; a tensor generator configured to encode the data into tensors across patients or patient cohorts, time, and treatment regimens; a synthetic model generator configured to use a machine learning process and said tensors to generate a synthetic model of a patient or a cohort of patients; a predicting means configured to predict an ADAS-Cog score for a patient or a cohort of patients under a target treatment regimen of the plurality of treatment regimens using the synthetic model; A system comprising:
18. the system processes data for a clinical trial to treat Alzheimer's disease; the data acquisition means is configured to collect data from at least two patient cohorts, a first patient cohort including a first subset of patients receiving a first treatment regimen and a second subset of patients receiving a second treatment regimen, and a second patient cohort including the first subset of patients receiving said first treatment regimen and a second subset of patients receiving a third treatment regimen, and the data includes information regarding longitudinal trajectories of ADAS-Cog scores of the patient cohorts over time; the tensor generation means is configured to encode the data into tensors across patient cohorts, time, and treatment plans; the synthetic model generation means is configured to use a machine learning process and the tensors to generate a first synthetic model for the first patient cohort and a second synthetic model for the second patient cohort; the prediction means is configured to predict an ADAS-Cog score for the first patient cohort under the third treatment regimen using the first composite model, and to predict an ADAS-Cog score for the second patient cohort under the second treatment regimen using the second composite model; 20. The system of claim 17.
19. A non-transitory computer readable storage medium comprising machine executable instructions which, when executed on a processor, cause the processor to perform the method of any one of claims 1 to 16.
