System and method of deploying neural network in optimizing vein characteristic during apheresis

Abstract

A system for optimizing fluid flow rate in a vein of a subject during apheresis procedure equipped with neural network training capacities. The system receives input of parametric vein characteristic of the subject and outputs a subject-specific flow rate continuously throughout the procedure to enable optimal control of the subject's vein condition while maintaining maximum flow rates specific to the subject.

Description

Attorney Docket No. F-7029 PCT (9362-0825)SYSTEM AND METHOD OF DEPLOYING NEURAL NETWORK IN OPTIMIZING VEIN CHARACTERISTIC DURING APHERESISThe present application claims the benefit of and priority to U.S. Provisional Application 63 / 760,282, filed February 19, 2025, which is incorporated herein by reference.Field of the Disclosure

[0001] The present subject matter relates to systems and methods for optimizing the fluid flow rate in a vein of a subject during an apheresis procedure. More particularly, the present subject matter relates to neural network systems for monitoring the vein condition of a donor / subject and providing an optimal donor-specific flow rate.Description of Related Art

[0002] In an apheresis procedure, whole blood is withdrawn from a donor / subject and is separated into its components (red blood cells, platelets, plasma and leukocytes). Certain components may be collected for further use or processing while other components may be returned to the donor. The separation of whole blood is typically accomplished in an automated procedure by centrifugation or membrane filtration.

[0003] Apheresis procedures, whether for donors or patients, require a robust vein monitoring system that can detect pressure changes indicative of potential vein issues, such as collapse or infiltration. Existing apheresis systems utilize pressure threshold algorithms designed to maintain vein pressure within a safe range i.e. , preventing it from dropping below a specified negative pressure during blood draw or exceeding a set positive pressure during blood reinfusion. When these thresholds are breached, the algorithms adjust the flow rate, adjusting either the draw or reinfusion, to mitigate the risk of vein collapse or infiltration.

[0004] While effective to a degree, current algorithms are limited by their reliance on a constant initial flow rate for all donors, which often results in slower than desired flow rates when pressure complications occur. These systems do not account for individual donor characteristics that can influence vein conditions and achievable flow rates.Attorney Docket No. F-7029 PCT (9362-0825) Furthermore, existing systems lack predictive or adaptive capabilities. The present disclosure is directed to developing donor-specific methods designed to enhance vein control systems.SUMMARY

[0005] There are several aspects of the present subject matter which may be embodied separately or together in the devices and systems described and claimed below. These aspects may be employed alone or in combination with other aspects of the subject matter described herein, and the description of these aspects together is not intended to preclude the use of these aspects separately or the claiming of such aspects separately or in different combinations as set forth in the claims appended hereto.

[0006] In accordance with a first aspect of the subject matter disclosed herein, a system for optimizing fluid flow rate in a vein of a subject during an apheresis procedure is provided. The system comprises an apheresis device comprising a reusable hardware component and a disposable fluid circuit mounted on the hardware component, the fluid circuit including a vein access device for accessing the vein of a subject. The reusable hardware component includes a user interface, one or more processors and a non-transitory memory. The one or more processors is / are configured for receiving a set of input parameters, indicative of a subject’s vein condition, in an input layer of a neural network. The neural network implemented in the one or more processors comprises an output layer interconnected to the input layer via a set of intermediate layers. The system is further configured to apply an activation function to a weighted sum of the set of the input parameter data received at one neuron in the set of intermediate layers, wherein the activation function interprets the weighted sum to produce a predicted vein characteristic; select, in response to a data qualification threshold being reached, a subset of the input layer while deactivating a remainder of another subset of the input layer from the at least one intermediate layer, the set of intermediate layers being configured in accordance with an initial matrix of weights. The system trains the neural network in accordance with the subset of input layers based at least in part upon adjusting the initial matrix of weights in accordanceAttorney Docket No. F-7029 PCT (9362-0825) with a supervised classification to predict, via the output layer, the at least one vein characteristic. The system deploys the trained neural network to provide, via the output layer, the predicted at least one vein characteristic as desired subject-specific vein characteristic based on the comparison.

[0007] In accordance with another aspect of the subject matter disclosed herein, a system for optimizing fluid flow rate in a vein of a subject during apheresis procedure is provided. The system comprises an apheresis device including a reusable hardware component and a disposable fluid circuit mounted on the hardware component. The fluid circuit includes a vein access device for accessing the vein of a subject. The reusable hardware component includes one or more processing units configured for connecting with a server device hosting a neural network, the server device including one or more processors, a non-transitory memory storing a set of instructions, wherein the one or more processors are configured for: receiving, from the apheresis device, a set of input parameters indicative of subject’s vein condition in an input layer of the neural network. The neural network implemented in the one or more processors and including an output layer interconnected to the input layer via a set of intermediate layers. Each of the set of input parameters is associated with a set of input parameter data of the subject’s vein condition. The system is further configured to apply an activation function to a weighted sum of the set of the input parameter data received at the at least one neuron in the set of intermediate layers, wherein the activation function interprets the weighted sum to produce the predicted at least one vein characteristic. The system is further configured to select, in response to data qualification threshold being reached, a subset of the input layer while deactivating a remainder of another subset of input layer from the at least one intermediate layer, the set of intermediate layers being configured in accordance with an initial matrix of weights. The system is also configured to train the neural network in accordance with the subset of input layers based at least in part upon adjusting the initial matrix of weights in accordance with a supervised classification to predict, via the output layer, the at least one vein characteristic. Finally, the system is configured to deploy the trained neural network to provide, via the output layer, the predicted at least one vein characteristic as desired subject-specific vein characteristic based on the comparison.Attorney Docket No. F-7029 PCT (9362-0825)

[0008] In accordance with a further aspect of the invention, provided is a method for optimizing fluid flow rate in a vein of a subject during apheresis procedure, which includes receiving, in an input layer of the neural network, from the apheresis device, a set of input parameters indicative of subject’s vein condition, the neural network implemented in the one or more processors and including an output layer interconnected to the input layer via a set of intermediate layers, wherein each of the set of input parameters is associated with a set of input parameter data of the subject’s vein condition. The method further includes applying an activation function to a weighted sum of the set of the input parameter data received at one neuron in the set of intermediate layers, wherein the activation function interprets the weighted sum to produce the predicted at least one vein characteristic. The method also includes selecting, in response to a data qualification threshold being reached, a subset of the input layer while deactivating a remainder of another subset of the input layer from the at least one intermediate layer, the set of intermediate layers being configured in accordance with an initial matrix of weights. The method includes training the neural network in accordance with the subset of input layers based at least in part upon adjusting the initial matrix of weights in accordance with a supervised classification to predict, via the output layer, the at least one vein characteristic. Finally, the method includes deploying the trained neural network to provide, via the output layer, the predicted at least one vein characteristic as desired subject-specific vein characteristic based on the comparison.BRIEF DESCRIPTION OF DRAWINGS

[0009] Fig. 1 illustrates an exemplary embodiment of a network system machine learning (ML) based neural network training for optimizing a donor-specific flow rate.

[0010] Fig. 2 illustrates an exemplary embodiment of an architecture of a network server computing system for ML based neural network training for optimizing a donorspecific flow rate.

[0011] Fig. 3A illustrates machine learning model diagram of an exemplary embodiment of system for neural network training for optimizing a donor-specific flow rate.Attorney Docket No. F-7029 PCT (9362-0825)

[0012] Fig. 3B illustrates a structure of a neuron in the exemplary neural network as shown in Fig. 3A, showing a weighted connection associated with the input parameter data and propagated forward to the output.

[0013] Fig. 4 illustrates an exemplary embodiment of a method of operation of a system for neural network training for optimizing a donor-specific flow rate.DETAILED DESCRIPTION

[0014] The embodiments disclosed herein are for the purpose of providing a description of the present subject matter, and it is understood that the subject matter may be embodied in various other forms and combinations not shown in detail.Therefore, specific designs and features disclosed herein are not to be interpreted as limiting the subject matter as defined in the accompanying claims.

[0015] The term “dynamic” and “dynamically” as used herein refers to actions performed during execution of the neural network training application.System Description

[0016] Fig. 1 illustrates, in an embodiment of a system 10 including an apheresis device 102 that includes or is otherwise linked to a module for training a neural network to optimize a fluid flow rate in a vein of a subject during an apheresis procedure. The apheresis device 102, is generally configured in accordance with the system described in U.S. Patent No. 12,083,258 to Patel et al. and the design of the Aurora® plasmapheresis machine marketed by Fenwal, Inc., of Lake Zurich, Illinois, an affiliate of Fresenius Kabi AG of Bad Homburg, Germany, and incorporated herein by reference. To withdraw blood from a donor / subject, separate blood into its components, collect a desired blood component and return the remaining components to the donor / subject. In one embodiment, apheresis device 102 is configured to collect an optimized volume of plasma.

[0017] As described in U.S. Patent No. 12,083,258, apheresis device 102 may include a reusable hardware component and a disposable fluid circuit for mounting onto the hardware unit. The disposable set includes a venipuncture needle 36, through which whole blood is drawn from the donor.

[0018] The hardware component includes or is otherwise linked to a programmable controller (control system) 102 and touch screen with a graphical user interface (“GUI”)Attorney Docket No. F-7029 PCT (9362-0825) through which the operator controls the procedure. For example, the GUI permits entry of any of donor and apheresis procedure parameters such as donor sex, donor height, donor weight, donor age, donor hematocrit / hemoglobin; a target saline infusion volume, and a target collection volume, vein pressure, vein pressure rate, donor flow rate which are described below. The touch screen also enables the operator to gather status information and handle error conditions.

[0019] In accordance with the present disclosure, the controller includes computer processor-executable instructions stored within a non-transitory memory. Server 100 is in communication via communication network 104 with apheresis device 102 for performing apheresis. Database 103, for example storing healthcare data accessible to medical data source software application 106 a, 106 b under execution, is communicatively accessible to server 100, and also to apheresis device 102.

[0020] Medical data source software applications 106 in another embodiment may be a web-based application program that executes on apheresis device 102. The apheresis device 102 can be used to access or acquire subject / donor medical data , for instance, by medical professional staff at medical clinics and hospitals.

[0021] Fig. 2 illustrates an embodiment of the architecture of a system 200 for training a machine learning neural network in monitoring vein condition and optimizing a fluid flow rate in a vein of a subject during an apheresis procedure. In this exemplary system, only the case in which the training system 200 is integrated within the server 100 is described. However, it should be appreciated that the configuration described herein is equally applicable to other architectural designs, such as integrating the training system 200 within the apheresis device 102. In another implementation, the training system 200 and the apheresis device 10 in the disintegration design may be located at the same facility center. Server computing system or device 100, also referred to herein as server, as illustrated in Fig. 1 are correspondingly represented in Fig. 2. In Fig. 2, the numeral ‘100’ is added to designate the training system 200, which may include, but is not limited to, the following active components: processor 201 , memory 202, display screen 203, input mechanisms 206 such as a keyboard or software-implemented touchscreen input functionality, and communication interface 208 for communicating via communication network 204. Memory 202 may be any type ofAttorney Docket No. F-7029 PCT (9362-0825) non-transitory system memory, storing instructions that are executable in processor 201 , including such as a static random access memory (SRAM), dynamic random access memory (DRAM), synchronous DRAM (SDRAM), read-only memory (ROM), or a combination thereof.

[0022] Neural network training logic module 210 includes processor-executable instructions stored in memory 202 of an internal control system in the apheresis device 102 or an external server 100, the instructions being executable in processor 201.Neural network training logic module 210 may include portions or sub-modules including input features module 220, layer selecting module 230, weights adjusting module 240 and output module 250 with a comparator 252 integrated therewith.

[0023] As illustrated in Figs. 2 and 3, processor 201 uses executable instructions of input features module 220 to receive input parameters indicative of a donor’s (or subject’s) vein condition at respective ones of a set of input layers 320 of the neural network 310. In an embodiment, the neural network 310 is implemented in one or more processors and comprises an output layer 350 interconnected to the set of input layers 320 via a set of intermediate (or hidden) layers 330, 340, each of the set of input parameters 322 being associated with input parameter data of the subject’s (donor’s) vein condition.

[0024] In an exemplary system, the input parameter (or feature) data 322 may include, but is not limited to, at least one of the following: vein pressure, vein pressure rate, donor flow rate (instantaneous blood flow to or from the donor), collection type, donor-specific characteristics (such as age, gender, height, weight, pulse, hematocrit, etc.), donor’s historical data record, or any combination thereof. In an exemplary system, the input parameter may include vein pressure 322a, vein pressure rate 322b and donor flow rate 322c. Vein pressure refers to the real-time blood pressure in the donor's vein, measured by sensors (not shown) in the apheresis device. The vein pressure rate represents the rate of change in vein pressure over a predefined time interval. The donor flow rate may be a commanded value or a measured value (from scales or other feedback sensors). Collection type refers to the type of fluid to be collected, such as plasma, platelets, red blood cells (RBCs). The donor’s historical data record may include any of the above parametric measurement data and any occurrenceAttorney Docket No. F-7029 PCT (9362-0825) of vascular complications such as vein collapse or hematoma from the previous blood collection procedure.

[0025] In an embodiment, initialization of the system may begin by using a primary model that considers the collection type, donor-specific characteristics, and historical donor data to determine an initial flow rate. Once the system is initialized, it then applies a secondary model that incorporates vein pressure, vein pressure rate, and the donor's instantaneous flow rate to determine a predicted vein characteristic, such as a desired donor flow rate as exemplified. It then continues with subsequent model training, leveraging input feature data, to optimize the flow rate to or from the donor throughout the process.

[0026] Processor 201 uses executable instructions stored in a layer selecting module 230 to select, in response to a data qualification threshold level (described below) being reached, a subset (i.e. neuron / node for input parameters) of the set of input layers 320 while deactivating from the intermediate layers 330, a remainder of subset of the set of the input layer 320, as illustrated in Fig. 3. In an embodiment, the set of intermediate / hidden layers 330, 340 are configured in accordance with an initial matrix of weights.

[0027] The learning process involves applying an activation function to a weighted sum of the set of input parameter data 320 received at a neuron in the set of intermediate layers 330, 340 of a neural network 310. In this context, the activation function serves to interpret the weighted sum of input data to produce an output that corresponds to the predicted vein characteristic, such as flow rate, pressure rate. The activation function performs this interpretation by applying a non-linear transformation to the weighted sum of inputs, allowing the neural network to model complex relationships in the data.

[0028] Examples of commonly used activation functions include Sigmoid, tanh (hyperbolic tangent), ReLU (Rectified Linear Unit), Leaky ReLU, Maxout, and ELU (Exponential Linear Unit). Each of these functions has its own unique properties that affect how the model learns and how effectively it captures non-linear patterns in the data.

[0029] Next, the data qualification threshold is applied, where the model checks whether the received input parametric data meet certain predefined criteria. ThisAttorney Docket No. F-7029 PCT (9362-0825) threshold is defined as the process of applying an activation function (denoted as / (z)) to a weighted sum of input feature data. The weighted sum can be expressed mathematically as z= valnwn, where n is the number of input data, valnare the input features and wnare the weights. Optionally, in another algorithm, a bias factor b may be added to the weighted sum (i.e. z= valnivn+h) allowing the activation function to model data set having complex relationship. Once the weighted sum is computed, the activation function interprets it to generate a prediction output, which is the predicted vein characteristic, such as flow rate, as shown in Fig. 3A.

[0030] For instance, as illustrated in Fig. 3B, the exemplary neural network shown in Fig. 3A consists of two hidden layers which further consist of nodes (or commonly called neurons), for example 4 neurons per layer, which represent the basic building blocks of the neural network. Each neuron acts as a function, processing input parameter data Vain to produce a single output of the weighted sum shown above. Alternatively, the neuron may act as a function (F), where Output = F(vah*wi, val2*W2, val3*ws)*W4 as shown in Fig. 3B. Each neuron produces a single output that is further passed on to other neurons in the network. Neurons are defined by an “in-degree” and an “out-degree”. The in-degree of a neuron refers to the number of inputs that the neuron receives whereas the out-degree represents the number of subsequent neurons to which the neuron sends its output to. The in-degree of a neuron in a current layer (e.g. layer 2) is equal to the out-degree of a neuron in the previous layer (e.g. layer 1) because a neuron in one layer connects to every neuron in the subsequent layer, and the number of connections between layers defines the structure of the network.

[0031] As mentioned, the purpose of having activation function is to introduce nonlinearity into the model and allow the model to learn complex non-linear relationship within the data. In an exemplified algorithm, the activation function may be ReLU function which yields:

[0032] Output = max(0, valfwi, val2*W2, val3*ws)*W4).

[0033] In response to the data qualification threshold being reached, the system selects a specific subset of neurons within the input layer 320 to become active, allowing these selected neurons 322 to forward the information through and within network 310. This selection process is performed by adjusting the weights and biasesAttorney Docket No. F-7029 PCT (9362-0825) according to the model's configuration based on the data validation / qualification criteria, which ensure that only the most relevant or qualified neurons 322 contribute to the learning process. Meanwhile, the remaining neurons 322 in the input layer 320 are deactivated, meaning they do not transmit their data to the next layer (i.e., the hidden or intermediate layers 330, 340). This selective activation of neurons reduces unnecessary computations and allows the network to focus on the most pertinent input features. The hidden intermediate layer 330, 340 is configured with respect to an initial matrix of weights, which is typically initialized with random values.

[0034] The system dynamically adjusts the activation of specific neurons within the network based on a predefined data qualification threshold, leading to an optimal selection of the most relevant features (input parameters) that contribute to accurate predictions of vein characteristics such as flow rate as described. The network's structure, defined by the matrix of weights, enables the efficient and accurate processing of input data through multiple layers, culminating in a predicted output.

[0035] The hidden or intermediate layers 330, 340, which are responsible for processing the input data, are designed and organized according to an initial matrix of weights. This matrix defines the connections between the input layer 320 and the intermediate layers 330, 340, guiding how information flows through the network 310. The weight matrix is central to the learning process, as it is gradually adjusted during training to optimize the model’s predictions and improve its performance over time.

[0036] The data qualification threshold is associated with one or more factors, including the category of the data source (e.g. blood and / or vein characteristics), the time proximity of data collection (e.g. recent or aged), the density of the data, and the confidence level tied to the input feature data provided to the input layers. In one embodiment, data density is a measure of the quality and sufficiency of the input data to ensure the neural network is trained effectively, enabling it to reliably and accurately optimize the desired vein characteristic — such as flow rate — when deployed.

[0037] Accordingly, based on the data source category, time proximity of data collection, data density, and confidence level of the input feature data, only input layers with sufficiently high-quality data are selected for activation in conjunction with one or more intermediate layers to create the trained neural network. Input layers with dataAttorney Docket No. F-7029 PCT (9362-0825) density or other attributes below the threshold are deactivated from neural network computations. This approach minimizes the consumption of computational resources, such as processing power and memory, for data attributes that do not meet the required threshold.

[0038] In a further embodiment, as the input data attributes rise to the level of a respective threshold requirement, the respective input layers 320 which were previously de-activated can now be activated via being dynamically linked to establish active nodes of the intermediate layers 330, 340 for computing operations in the neural network 310, and are thus re-engaged in the neural network training. In still further embodiment, with regard to the time proximity of data collection, more recent data can be prioritized, or weighted more heavily, than older data. With regard to the data source category, data collected in more controlled or monitored contexts by clinical staff may be weighted more heavily, or accorded a higher quality status, than more casually collected donor vein condition data.

[0039] Still referring to Figs. 2 and 3A, processor 201 uses executable instructions stored in the weights adjusting module 240 to train the neural network 310 in accordance with the subset of input layers 320 based at least in part upon adjusting the initial matrix of weights using a supervised classification that provides, via the output layer 350, prediction of the at least one view characteristic, such as flow rate as exemplified.

[0040] In one embodiment, training the neural network classifier includes adjusting the initial matrix of weights by backpropagation to diminish an error matrix between the predicted at least one vein characteristic generated at the output layer and the actual measurement of the characteristic in accordance with the supervised classification. The adjusting, in some embodiments, comprises recursively adjusting the initial matrix of weights by backpropagation 360 in diminishment of the error matrix computed at the output layer 350 of the neural network classifier. The error matrix may be computed by applying loss function such as mean squared error (MSE). The value of the error matrix is derived from a gradient of the loss function, MSE as exemplified, with respect to each weight associated with the input feature data using chain rule.Attorney Docket No. F-7029 PCT (9362-0825)

[0041] To reduce the error, the training model may be implemented in accordance with executable instructions stored in a weight adjusting module 240. The neural network, in one embodiment, is configured with a set of input layers 320, an output layer 350, and one or more intermediate layers 330, 340 connecting the input 320 and output 350 layers. In embodiments, the input layers 320 are associated with input features that relate to parameter data of a donor, such as, but not limited to, data sourced or accessed via the apheresis device 102 and / or database 103 connected with the neural network module 110.

[0042] The supervised classification comprises comparing the predicted at least one vein characteristic, flow rate as exemplified, by a trained neural network, with a predefined acceptance criterion-for example, this criterion represents an acceptable flow rate value within a predefined threshold range of one or more another vein characteristic, such as pressure rate. As shown in Fig. 3A, a comparator 350 serves as a form of quality control to ensure that the predicted flow rate output of the neural network falls within an acceptable value to comply with the safety guideline.Example

[0043] To meet the acceptance criteria, an allowable donor return flow rate 120 ml / min may occur within a pressure range of 100-150 mmHg. If an exemplary model outputs a prediction of 140 ml / min while the pressure is within the above-specified range, an error of 20 ml / min is fed back to the module during backpropagation training. This error informs the neural network on how to adjust the weights, commonly using a loss function like MSE. The MSE is minimized during training to improve the accuracy of the model. It is common to train a model using 80% of a dataset and then use the remaining 20% of the dataset to test the accuracy of the model.

[0044] The term supervised classification as used herein refers to a supervised machine learning process which includes having a fully labeled set of data constituting parametric acceptance while training the machine learning algorithm. Fully labeled means that each example in the training dataset is tagged with the parametric acceptance that the algorithm, once trained as the matrix of weights is adjusted, should derive on its own.Attorney Docket No. F-7029 PCT (9362-0825)

[0045] In an embodiment of the present system, the neural network 310 is deployed as a trained neural network, to evaluate the input parameter data and predict the vein characteristic of interest (flow rate). Typically, neural network 310 has undergone training on a diverse dataset, allowing it to identify patterns and relationships between various input features and the target vein characteristic (flow rate). A predicted vein characteristic, which represents its assessment based on the inputs, is produced at the output layer 350. The predicted characteristic is subsequently analyzed and compared against the acceptance criteria to evaluate its integrity and suitability for application in the donor-specific apheresis procedure. Based on the comparison between the predicted characteristic and the predefined threshold, the neural network confirms the predicted characteristic as the final flow rate to be applied in the apheresis procedure.

[0046] The neural networks described herein in the embodiments refer to an artificial intelligence (Al) based neural network, including machine learning (ML) or deep learning (DL) models.Method

[0047] FIG. 4 illustrates a method of operation, in one embodiment 400, of a system for neural network training of patient diagnostic states. Method 400 embodiment depicted is performed by one or more processors 201 of a control system 102 integrated in an apheresis device or a server computing device 101. In describing and performing the embodiments of FIG. 4, the examples of FIG. 1 and FIG. 2 are incorporated for purposes of illustrating suitable components or elements for performing a step or sub-step being described.

[0048] Examples of method steps described herein relate to the use of server 100 for implementing the techniques described. According to one embodiment, the techniques are performed by neural network training logic module 110 of server 100 in response to the processor 201 executing one or more sequences of software logic instructions that constitute neural network training logic module 110.

[0049] In an embodiment, neural network training logic module 110, 210 may include one or more sequences of instructions within sub-modules including input features module 220, layer selecting module 230 and the weight adjusting module 240. Such instructions may be read into memory 202 from machine-readable medium, such asAttorney Docket No. F-7029 PCT (9362-0825) memory storage devices. In executing the sequences of instructions contained in input features module 220, layer selecting module 230 and weights adjusting module 240 of neural network training logic module 210 in memory 202, output module 250 connected to or incorporated into neural network training logic module 210, and the processor 201 performs the process steps described herein. In alternative implementations, at least some hard-wired circuitry may be used in place of, or in combination with, the software logic instructions to implement examples described herein. Thus, the examples described herein are not limited to any particular combination of hardware circuitry and software instructions.

[0050] As illustrated in Figs. 2-4, in an exemplary method at step 410, processor 201 executes instructions of input features module 210 to receive a set of input parameters indicative of a donor’s (or subject’s) vein condition at respective points of a set of input layers 320 of the neural network 310. In an embodiment, the neural network 310 is implemented in one or more processors and includes an output layer 350 interconnected to the set of input layers 320 via a set of intermediate layers 330, 340, each of the set of input parameters being associated with input parameter data of the donor’s vein condition.

[0051] In the method, the input parameter (or feature) data may encompass, but is not limited to include at least one of the following: vein pressure, vein pressure rate, donor flow rate (instantaneous blood flow to or from the donor), collection type, donorspecific characteristics (such as age, gender, height, weight, pulse, hematocrit, etc.), donor’s historical data record, or any combination thereof. In an exemplary method, the input parameters may include vein pressure 322a, vein pressure rate 322b and donor flow rate 322c. Vein pressure refers to the real-time blood pressure in the donor's vein, measured by sensors (not shown) in the apheresis device 102. The vein pressure rate represents the rate of change in vein pressure over a predefined time interval. The donor flow rate may be a commanded value or a measured value (from scales or other feedback sensors). Collection type refers to the type of fluid to be collected, such as plasma, platelets, red blood cells (RBCs). Donor’s historical data record may include any of the above parametric measurement data and any occurrence of vascularAttorney Docket No. F-7029 PCT (9362-0825) complications such as vein collapse or hematoma from the previous blood collection procedure.

[0052] In another embodiment, initialization of the system may begin by using a primary model that considers the collection type, donor-specific characteristics, and historical donor data to determine an initial flow rate. Once the system is initialized, it then applies a secondary model that incorporates vein pressure, vein pressure rate, and the donor's instantaneous flow rate to determine a predicted vein characteristic, such as a desired donor flow rate as exemplified. It then continues with subsequent model training, leveraging input feature data, to optimize the flow rate to or from the donor throughout the process.

[0053] At steps 420 and 430 of Fig. 4, processor 201 executes instructions included in layer selecting module 230, 330 to (i) apply an activation function to a weighted sum of the set of input parameter data received at a neuron in the set of intermediate layers of a neural network, and (ii) select, in responsive to a data qualification threshold level being reached, a subset (i.e. neuron / node for input parameters) of the set of input layers, while deactivating from the intermediate layers a remainder of another subset of the input layer 320, as illustrated in Fig. 3A. As exemplified, the set of intermediate I hidden layers 330, 340 are configured in accordance with an initial matrix of weights. In this context, the activation function serves to interpret the weighted sum of input data to produce an output that corresponds to the predicted vein characteristic, such as flow rate, pressure rate. The activation function performs this interpretation by applying a non-linear transformation to the weighted sum of inputs, allowing the neural network to model complex relationships in the data.

[0054] Examples of commonly used activation functions include Sigmoid, tanh (hyperbolic tangent), ReLU (Rectified Linear Unit), Leaky ReLU, Maxout, and ELU (Exponential Linear Unit). Each of these functions has its own unique properties that affect how the model learns and how effectively it captures non-linear patterns in the data.

[0055] Next, a data qualification threshold may be applied, where the model checks whether the received input parametric data meet certain predefined criteria. This threshold is defined as the process of applying an activation function (denoted as / (z))Attorney Docket No. F-7029 PCT (9362-0825) to a weighted sum of input feature data. The weighted sum can be expressed mathematically as z= valnwn, where n is the number of input feature data, valnare the input feature data and wnare the weights . Optionally, in another algorithm, a bias term b may be added to the weighted sum (i.e. z= valnwn+&) allowing the activation function to model data set having complex relationship. Once the weighted sum is computed, the activation function interprets it to generate a prediction output, which is the predicted vein characteristic, such as flow rate, as shown in Fig. 3A.

[0056] For instance, as illustrated in Fig. 3B, an exemplary neural network shown in Fig. 3A consists of two hidden layers which further consist of nodes (or commonly called neurons), for example 4 neurons per layer, which represent the basic building blocks of the neural network. Each neuron acts as a function, processing input parameter data VaL to produce a single output of the weighted sum shown above. Alternatively, the neuron may act as a function (F), where Output = F(vah*wi, val2*w2, val3*w3)*w4as shown in Fig. 3B. Each neuron produces a single output that is further passed on to other neurons in the network. Neurons are defined by an “in-degree” and an “out-degree”. The in-degree of a neuron refers to the number of inputs that the neuron receives whereas the out-degree represents the number of subsequent neurons to which the neuron sends its output. The in-degree of a neuron in a current layer (e.g. layer 2) is equal to the out-degree of a neuron in the previous layer (e.g. layer 1) because a neuron in one layer connects to every neuron in the subsequent layer, and the number of connections between layers defines the structure of the network.

[0057] As mentioned, the purpose of having activation function is to introduce nonlinearity into the model and allow the model to learn complex non-linear relationship within the data. In an exemplified algorithm, the activation function may be ReLLI function which yields:

[0058] Output = max(0, vah*wi, val2*w2, val3*w3)*w4).

[0059] In response to the data qualification threshold being reached, the system selects a specific subset of neurons within the input layer 320 to become active, allowing these selected neurons 322 to pass their information forward through the network 310. This selection process is performed by adjusting the weights or together with the biases according to the model's configuration based on the data validation / Attorney Docket No. F-7029 PCT (9362-0825) qualification criteria, which ensure that only the most relevant or qualified neurons 322 contribute to the learning process. Meanwhile, the remaining neurons 322 in the input layer 320 are deactivated, meaning they do not transmit their data to the next layer (i.e., the hidden or intermediate layers 330, 340). This selective activation of neurons reduces unnecessary computations and allows the network to focus on the most pertinent input features. The hidden intermediate layers 330, 340 are configured with respect to an initial matrix of weights, which is typically initialized with random values.

[0060] The system dynamically adjusts the activation of specific neurons within the network based on a predefined data qualification threshold, leading to an optimal selection of the most relevant features (input parameters) that contribute to accurate predictions of vein characteristics such as flow rate as exemplified. The network's structure, defined by the matrix of weights, enables the efficient and accurate processing of input data through multiple layers, culminating in the predicted output.

[0061] The hidden or intermediate layers 330, 340, which are responsible for processing the input data, are designed and organized according to an initial matrix of weights. This matrix defines the connections between the input layer 320 and the intermediate layers 330, 340, guiding how information flows through the network 310. The weight matrix is central to the learning process, as it is gradually adjusted during training to optimize the model's predictions and improve its performance over time.

[0062] The data qualification threshold is associated with one or more factors, including the category of the data source, the time proximity of data collection, the density of the data, and the confidence level tied to the input feature data provided to the input layers. In one embodiment, data density is a measure of the quality and sufficiency of the input data to ensure the neural network is trained effectively, enabling it to reliably and accurately optimize the desired vein characteristic — such as flow rate — when deployed.

[0063] Accordingly, based on the data source category, time proximity of data collection, data density, and confidence level of the input feature data, only input layers with sufficiently high-quality data are selected for activation in conjunction with one or more intermediate layers to create the trained neural network. Input layers with data density or other attributes below the threshold are deactivated from neural networkAttorney Docket No. F-7029 PCT (9362-0825) computations. This approach minimizes the consumption of computational resources, such as processing power and memory, for data attributes that do not meet the required threshold.

[0064] In a further embodiment, as the input data attributes rise to the level of a respective threshold requirement, the respective input layers 320 which were previously de-activated can now be activated via being dynamically linked to establish active nodes of the intermediate layers 330, 340 for computing operations in the neural network 310, and are thus re-engaged in the neural network training. In a still further embodiment, with regard to the time proximity of data collection, more recent data can be prioritized, or weighted more heavily, than older data. With regard to the data source category, data collected in more controlled or monitored contexts by clinical staff may be weighted more heavily, or accorded a higher quality status, than more casually collected donor vein condition data.

[0065] At step 440, processor 201 executes instructions included in weights adjusting module 240 to train the neural network 310 in accordance with the subset of the input layers 320 based at least in part upon adjusting the initial matrix of weights using a supervised classification that provides, via the output layer 350, prediction of the at least one view characteristic, such as flow rate as exemplified.

[0066] In still another embodiment, training the neural network classifier comprises adjusting the initial matrix of weights by backpropagation 360 to diminish an error matrix between the predicted at least one vein characteristic generated at the output layer 350 and the actual measurement of the characteristic in accordance with the supervised classification. The adjusting, in some embodiments comprises recursively adjusting the initial matrix of weights by backpropagation 360 in diminishment of the error matrix computed at the output layer 350 of the neural network classifier. The error matrix may be computed by applying loss function such as mean squared error (MSE). The value of the error matrix is derived from a gradient of the loss function, MSE as exemplified, with respect to each weight associated with the input feature data using chain rule.

[0067] To reduce the error, the training model may be implemented in accordance with executable instructions stored in the weights adjusting module 240. The neural network, in one embodiment, is configured with a set of input layers 320, an output layerAttorney Docket No. F-7029 PCT (9362-0825) 350, and one or more intermediate layers 330, 340 connecting the input 320 and output 350 layers. In an embodiment, the input layers 320 are associated with input features that relate to parameter data of a donor, such as, but not limited to, parametric data sourced or accessed via the apheresis device 102 and / or database 103 connected with the neural network module 110.

[0068] At step 450, processor 201 executes instructions included in output module 250 to compare the predicted at least one vein characteristic, flow rate as exemplified, by a trained neural network, with a predefined acceptance criterion-for example, this criterion represents an acceptable flow rate value within a predefined threshold range of one or more another vein characteristic, such as pressure rate. As shown in Fig. 3, the comparator 352 serves as a form of quality control to ensure that the predicted flow rate output of the neural network falls within an acceptable value to comply with the safety guideline.Example

[0069] To meet the acceptance criteria set out in the safety guideline, an allowable donor return flow rate 120 ml / min may occur within a pressure range of 100-150 mmHg. If an exemplary model outputs a prediction of 140 ml / min while the pressure is within the above-specified range, an error of 20 ml / min is fed back to the module during backpropagation training. This error informs the neural network on how to adjust the weights, commonly using a loss function like MSE. The MSE is minimized during training to improve the accuracy of the model. It is common to train a model using 80% of a dataset and then use the remaining 20% of the dataset to test the accuracy of the model.

[0070] The term supervised classification as used herein refers to a supervised machine learning process which includes having a fully labeled set of data constituting parametric acceptance while training the machine learning algorithm. Fully labeled means that each example in the training dataset is tagged with the parametric acceptance that the algorithm, once trained as the matrix of weights is adjusted, should derive on its own.

[0071] At step 460, processor 201 executes instruction included in output module 250 to deploy the trained neural network, to evaluate the input parameter data andAttorney Docket No. F-7029 PCT (9362-0825) predict the vein characteristic of interest (flow rate). The neural network 310 has undergone training on a diverse dataset, allowing it to identify patterns and relationships between various input features and the target vein characteristic (flow rate). A predicted vein characteristic, which represents its assessment based on the inputs, is produced at the output layer 350. The predicted characteristic is subsequently analyzed and compared against the acceptance criteria to evaluate its integrity and suitability for application in the donor-specific procedure. At step 460, based on the comparison between the predicted characteristic and the predefined threshold, the output module 250 in the neural network confirms the predicted characteristic as the final flow rate to be applied in the procedure.

[0072] It will be seen that the techniques described herein, among other advantages, enable training of a machine learning neural network that accurately optimizes a donorspecific flow rate during apheresis procedure. In particular, solutions provided herein allow machine learning neural networks to be feasible for deployment and use even for large data sets, as can be experienced in regard to machine learning applications for monitoring vein condition. Among other benefits, solutions herein enable neural network training in accordance with dynamically selected input layers of the neural network, providing reduced computational power, increased computational efficiency, robust vein monitor and improved system response times associated with vascular issues arising from apheresis.ASPECTS

[0073] Aspect 1. A system for optimizing fluid flow rate in a vein of a subject during apheresis procedure, comprising an apheresis device comprising a reusable hardware component and a disposable fluid circuit mounted on said hardware component, said fluid circuit including a vein access device for accessing the vein of a subject; said reusable hardware component including a user interface, one or more processors and a non-transitory memory storing, wherein said one or more processors is configured for: receiving, in an input layer of the neural network, from the apheresis device, a set of input parameters indicative of subject’s vein condition, the neural network implementedAttorney Docket No. F-7029 PCT (9362-0825) in the one or more processors and comprising an output layer interconnected to the input layer via a set of intermediate layer, each of the set of input parameters being associated with a set of input parameter data of the subject’s vein condition; applying an activation function to a weighted sum of the set of the input parameter data received at the at least one neuron in the set of intermediate layers, wherein the activation function interprets the weighted sum to produce the predicted at least one vein characteristic; selecting, in response to data qualification threshold being reached, a subset of the input layer while deactivating a remainder of another subset of the input layer from the at least one intermediate layer, the set of intermediate layers being configured in accordance with an initial matrix of weights; training the neural network in accordance with the subset of input layers based at least in part upon adjusting the initial matrix of weights in accordance with a supervised classification to predict, via the output layer, the at least one vein characteristic; and deploying the trained neural network to provide, via the output layer, the predicted at least one vein characteristic as desired subjectspecific vein characteristic based on the comparison.

[0074] Aspect 2. The system of Aspect 1 , wherein the adjusting comprises recursively adjusting the initial matrix of weights by backpropagation in minimizing error between the predicted at least one vein characteristic and the actual measurement of the characteristic.

[0075] Aspect 3. The system of Aspect 1 or 2 wherein said one or more processors is further configured for comparing the predicted at least one vein characteristic with an acceptable value within a predefined threshold range of one or more another vein characteristic.

[0076] Aspect 4. The system of Aspect 1 , wherein the set of input parameters includes at least one of vein pressure, vein pressure rate, user flow rate, collection type, subject-specific characteristics, and subject data record, wherein the subject-specific characteristics includes one of a subject age, gender, height, weight, pulse, hematocrit.

[0077] Aspect 5. The system of Aspect 1 , wherein the one or more processors further configured for initializing the neural network using a primary model to determine an initial at least one vein characteristic based on at least one subset of the set of input parameters.Attorney Docket No. F-7029 PCT (9362-0825)

[0078] Aspect 6. The system of Aspect 5, wherein the one or more processors further configured for training the neural network by applying a secondary model to optimize the at least one vein characteristic based on another at least one subset of the set of input parameters.

[0079] Aspect 7. The system of Aspect 1 , wherein the trained neural network comprises a first neural network training iteration and further comprising at least a second neural network training iteration that includes re-connecting, from a de-activated state, at least one of the remainder subset of the input layer to the intermediate layers, in responsive to a data qualification threshold being reached for the at least one of the remainder subset of the input layer.

[0080] Aspect 8. A system for optimizing fluid flow rate in a vein of a subject during apheresis procedure, comprising an apheresis device comprising a reusable hardware component and a disposable fluid circuit mounted on said hardware component, said fluid circuit including a vein access device for accessing the vein of a subject; said reusable hardware component including one or more processing units configured for connecting with a server device hosting a neural network; the server device comprising one or more processors, a non-transitory memory storing a set of instruction, wherein said one or more processors configured for: receiving, in an input layer of the neural network, from the apheresis device, a set of input parameters indicative of subject’s vein condition, the neural network implemented in the one or more processors and comprising an output layer interconnected to the input layer via a set of intermediate layer, each of the set of input parameters being associated with a set of input parameter data of the subject’s vein condition; applying an activation function to a weighted sum of the set of the input parameter data received at the at least one neuron in the set of intermediate layers, wherein the activation function interprets the weighted sum to produce the predicted at least one vein characteristic; selecting, in response to data qualification threshold being reached, a subset of the input layer while deactivating a remainder of another subset of the input layer from the at least one intermediate layer, the set of intermediate layers being configured in accordance with an initial matrix of weights; training the neural network in accordance with the subset of input layers based at least in part upon adjusting the initial matrix of weights in accordance with aAttorney Docket No. F-7029 PCT (9362-0825) supervised classification to predict, via the output layer, the at least one vein characteristic; and deploying the trained neural network to provide, via the output layer, the predicted at least one vein characteristic as desired subject-specific vein characteristic based on the comparison.

[0081] Aspect 9. The system of Aspect 8, wherein the adjusting comprises recursively adjusting the initial matrix of weights by backpropagation in minimizing error between the predicted at least one vein characteristic and the actual measurement of the characteristic.

[0082] Aspect 10. The system of Aspect 8 or 9 wherein said one or more processors is further configured for comparing the predicted at least one vein characteristic with an acceptable value within a predefined threshold range of one or more another vein characteristic.

[0083] Aspect 11. The system of Aspect 8 wherein the set of input parameters includes at least one of vein pressure, vein pressure rate, user flow rate, collection type, subject-specific characteristics, and subject data record, wherein the subject-specific characteristics includes one of a subject age, gender, height, weight, pulse, hematocrit.

[0084] Aspect 12. The system of Aspect 8, wherein the one or more processors further configured for initializing the neural network using a primary model to determine an initial at least one vein characteristic based on at least one subset of the set of input parameters.

[0085] Aspect 13. The system of Aspect 12, wherein the one or more processors further configured for training the neural network by applying a secondary model to optimize the at least one vein characteristic based on another at least one subset of the set of input parameters.

[0086] Aspect 14. The system of Aspect 8, wherein the trained neural network comprises a first neural network training iteration and further comprising at least a second neural network training iteration that includes re-connecting, from a de-activated state, at least one of the remainder subset of the input layer to the intermediate layers, in responsive to a data qualification threshold being reached for the at least one of the remainder subset of the input layer.Attorney Docket No. F-7029 PCT (9362-0825)

[0087] Aspect 15. A method for optimizing fluid flow rate in a vein of a subject during apheresis procedure, comprising: receiving, in an input layer of the neural network, from the apheresis device, a set of input parameters indicative of subject’s vein condition, the neural network implemented in the one or more processors and comprising an output layer interconnected to the input layer via a set of intermediate layer, each of the set of input parameters being associated with a set of input parameter data of the subject’s vein condition; applying an activation function to a weighted sum of the set of the input parameter data received at the at least one neuron in the set of intermediate layers, wherein the activation function interprets the weighted sum to produce the predicted at least one vein characteristic; selecting, in response to data qualification threshold being reached, a subset of the input layer while deactivating a remainder of another subset of the input layer from the at least one intermediate layer, the set of intermediate layers being configured in accordance with an initial matrix of weights; training the neural network in accordance with the subset of input layers based at least in part upon adjusting the initial matrix of weights in accordance with a supervised classification to predict, via the output layer, the at least one vein characteristic; and deploying the trained neural network to provide, via the output layer, the predicted at least one vein characteristic as desired subject-specific vein characteristic based on the comparison.

[0088] Aspect 16. The method of Aspect 15, wherein the adjusting comprises recursively adjusting the initial matrix of weights by backpropagation in minimizing error between the predicted at least one vein characteristic and the actual measurement of the characteristic.

[0089] Aspect 17. The method of Aspect 15 or 16 further comprises comparing the predicted at least one vein characteristic with an acceptable value within a predefined threshold range of one or more another vein characteristic.

[0090] Aspect 18. The method of Aspect 15, wherein the set of input parameters includes at least one of vein pressure, vein pressure rate, user flow rate, collection type, user-specific characteristics, and user data record, wherein the subject-specific characteristics includes one of a user age, gender, height, weight, pulse, hematocrit.Attorney Docket No. F-7029 PCT (9362-0825)

[0091] Aspect 19. The method of Aspect 15 further comprising initializing the neural network using a primary model to determine an initial at least one vein characteristic based on at least one subset of the set of input parameters.

[0092] Aspect 20. The method of Aspect 19 further comprising training the neural network by applying a secondary model to optimize the at least one vein characteristic based on another at least one subset of the set of input parameters.

[0093] Aspect 21. The method of Aspect 15 further comprising re-connecting, from a de-activated state, at least one of the remainder subset of the input layer to the intermediate layers, in response to a data qualification threshold being reached for the at least one of the remainder subset of the input layer.

[0094] It will be understood that the embodiments described above are illustrative of some of the applications of the principles of the present subject matter. Numerous modifications may be made by those skilled in the art without departing from the spirit and scope of the claimed subject matter, including those combinations of features that are individually disclosed or claimed herein. For these reasons, the scope hereof is not limited to the above description but is as set forth in the following claims, and it is understood that claims may be directed to the features hereof, including as combinations of features that are individually disclosed or claimed herein.

Claims

Attorney Docket No. F-7029 PCT (9362-0825) CLAIMS1. A system for optimizing fluid flow rate in a vein of a subject during apheresis procedure, comprising:an apheresis device comprising a reusable hardware component and a disposable fluid circuit mounted on said hardware component, said fluid circuit including a vein access device for accessing the vein of a subject;said reusable hardware component including a user interface, one or more processors and a non-transitory memory storing, wherein said one or more processors is configured for:receiving, in an input layer of the neural network, from the apheresis device, a set of input parameters indicative of subject’s vein condition, the neural network implemented in the one or more processors and comprising an output layer interconnected to the input layer via a set of intermediate layer, each of the set of input parameters being associated with a set of input parameter data of the subject’s vein condition;applying an activation function to a weighted sum of the set of the input parameter data received at the at least one neuron in the set of intermediate layers, wherein the activation function interprets the weighted sum to produce the predicted at least one vein characteristic;selecting, in response to data qualification threshold being reached, a subset of the input layer while deactivating a remainder of another subset of the input layer from the at least one intermediate layer, the set of intermediate layers being configured in accordance with an initial matrix of weights;training the neural network in accordance with the input layer based at least in part upon adjusting the initial matrix of weights in accordance with a supervised classification to predict, via the output layer, the at least one vein characteristic;anddeploying the trained neural network to provide, via the output layer, the predicted at least one vein characteristic as desired subject-specific vein characteristic based on the comparison.Attorney Docket No. F-7029 PCT (9362-0825)2. The system of claim 1 , wherein the adjusting comprises recursively adjusting the initial matrix of weights by backpropagation in minimizing error between the predicted at least one vein characteristic and the actual measurement of the characteristic.

3. The system of claim 1 or 2, wherein said one or more processors is further configured for comparing the predicted at least one vein characteristic with an acceptable value within a predefined threshold range of one or more another vein characteristic.

4. The system of claim 1 wherein the set of input parameters includes at least one of vein pressure, vein pressure rate, user flow rate, collection type, subject-specific characteristics, and subject data record, wherein the subject-specific characteristics includes one of a subject age, gender, height, weight, pulse, hematocrit.

5. The system of claim 1 , wherein the one or more processors further configured for initializing the neural network using a primary model to determine an initial at least one vein characteristic based on at least one subset of the set of input parameters.

6. The system of claim 5, wherein the one or more processors further configured for training the neural network by applying a secondary model to optimize the at least one vein characteristic based on another at least one subset of the set of input parameters.

7. The system of claim 1 , wherein the trained neural network comprises a first neural network training iteration and further comprising at least a second neural network training iteration that includes re-connecting, from a de-activated state, at least one of the remainder subset of the input layer to the intermediate layers, in responsive to a data qualification threshold being reached for the at least one of the remainder subset of the input layer.Attorney Docket No. F-7029 PCT (9362-0825) 8. A system for optimizing fluid flow rate in a vein of a subject during apheresis procedure, comprising:an apheresis device comprising a reusable hardware component and a disposable fluid circuit mounted on said hardware component, said fluid circuit including a vein access device for accessing the vein of a subject;said reusable hardware component including one or more processing units configured for connecting with a server device hosting a neural network;the server device comprising one or more processors, a non-transitory memory storing a set of instruction, wherein said one or more processors configured for:receiving, in an input layer of the neural network, from the apheresis device, a set of input parameters indicative of subject’s vein condition, the neural network implemented in the one or more processors and comprising an output layer interconnected to the input layer via a set of intermediate layer, each of the set of input parameters being associated with a set of input parameter data of the subject’s vein condition;applying an activation function to a weighted sum of the set of the input parameter data received at the at least one neuron in the set of intermediate layers, wherein the activation function interprets the weighted sum to produce the predicted at least one vein characteristic;selecting, in response to data qualification threshold being reached, a subset of the input layer while deactivating a remainder of another subset of the input layer from the at least one intermediate layer, the set of intermediate layers being configured in accordance with an initial matrix of weights;training the neural network in accordance with the input layer based at least in part upon adjusting the initial matrix of weights in accordance with a supervised classification to predict, via the output layer, the at least one vein characteristic; anddeploying the trained neural network to provide, via the output layer, the predicted at least one vein characteristic as desired subject-specific vein characteristic based on the comparison.Attorney Docket No. F-7029 PCT (9362-0825) 9. The system of claim 8, wherein the adjusting comprises recursively adjusting the initial matrix of weights by backpropagation in minimizing error between the predicted at least one vein characteristic and the actual measurement of the characteristic.

10. The system of claim 8 or 9, wherein said one or more processors is further configured for comparing the predicted at least one vein characteristic with an acceptable value within a predefined threshold range of one or more another vein characteristic.

11. The system of claim 8 wherein the set of input parameters includes at least one of vein pressure, vein pressure rate, user flow rate, collection type, subject-specific characteristics, and subject data record, wherein the subject-specific characteristics includes one of a subject age, gender, height, weight, pulse, hematocrit.

12. The system of claim 8, wherein the one or more processors further configured for initializing the neural network using a primary model to determine an initial at least one vein characteristic based on at least one subset of the set of input parameters.

13. The system of claim 12, wherein the one or more processors further configured for training the neural network by applying a secondary model to optimize the at least one vein characteristic based on another at least one subset of the set of input parameters.

14. The system of claim 8, wherein the trained neural network comprises a first neural network training iteration and further comprising at least a second neural network training iteration that includes re-connecting, from a de-activated state, at least one of the remainder subset of the input layer to the intermediate layers, in responsive to a data qualification threshold being reached for the at least one of the remainder subset of the input layer.Attorney Docket No. F-7029 PCT (9362-0825) 15. A method for optimizing fluid flow rate in a vein of a subject during apheresis procedure, comprising:receiving, in an input layer of the neural network, from the apheresis device, a set of input parameters indicative of subject’s vein condition, the neural network implemented in the one or more processors and comprising an output layer interconnected to the input layer via a set of intermediate layer, each of the set of input parameters being associated with a set of input parameter data of the subject’s vein condition;applying an activation function to a weighted sum of the set of the input parameter data received at the at least one neuron in the set of intermediate layers, wherein the activation function interprets the weighted sum to produce the predicted at least one vein characteristic;selecting, in response to data qualification threshold being reached, a subset of the input layer while deactivating a remainder of another subset of the input layer from the at least one intermediate layer, the set of intermediate layers being configured in accordance with an initial matrix of weights;training the neural network in accordance with the input layer based at least in part upon adjusting the initial matrix of weights in accordance with a supervised classification to predict, via the output layer, the at least one vein characteristic; anddeploying the trained neural network to provide, via the output layer, the predicted at least one vein characteristic as desired subject-specific vein characteristic based on the comparison.

16. The method of claim 15, wherein the adjusting comprises recursively adjusting the initial matrix of weights by backpropagation in minimizing error between the predicted at least one vein characteristic and the actual measurement of thecharacteristic.Attorney Docket No. F-7029 PCT (9362-0825) 17. The method of claim 15 or 16 further comprises comparing the predicted at least one vein characteristic with an acceptable value within a predefined threshold range of one or more another vein characteristic.

18. The method of claim 15 wherein the set of input parameters includes at least one of vein pressure, vein pressure rate, user flow rate, collection type, user-specific characteristics, and user data record, wherein the subject-specific characteristics includes one of a user age, gender, height, weight, pulse, hematocrit.

19. The method of claim 15 further comprising initializing the neural network using a primary model to determine an initial at least one vein characteristic based on at least one subset of the set of input parameters.

20. The method of claim 19 further comprising training the neural network by applying a secondary model to optimize the at least one vein characteristic based on another at least one subset of the set of input parameters.

21. The method of claim 15 further comprising re-connecting, from a de-activated state, at least one of the remainder subset of the input layer to the intermediate layers, in response to a data qualification threshold being reached for the at least one of the remainder subset of the input layer.

Images