Systems and methods for managing multidrug delivery using infusion pumps

Abstract

Systems and methods of multidrug infusion management perform and / or include, in an assembly including a plurality of infusion pumps; an infusion manifold including a plurality of fluid input ports respectively connected to the plurality of infusion pumps and including a fluid mixing region; and a controller: modeling, in real-time, a spatiotemporal evolution of a drug concentration in the fluid mixing region in accordance with a transport model, and dynamically adjusting a rate of fluid flow from at least one of the plurality of infusion pumps to at least one of the plurality of fluid input ports of the infusion manifold based on the modeling.

Description

Client No: MGH 2024-047-03 Quarles: 125141.04694 PCT PATENT APPLICATION FOR SYSTEMS AND METHODS FOR MANAGING MULTIDRUG DELIVERY USING INFUSION PUMPS Robert A. Peterfreund David E. Arney Andrea Detry Nathaniel M. Sims Samir Mitragotri Vineeth Chandran Suja QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) SYSTEMS AND METHODS FOR MANAGING MULTIDRUG DELIVERY USING INFUSION PUMPS CROSS-REFERENCE TO RELATED APPLICATIONS

[0001] This application claims the benefit of and priority to U.S. Provisional Patent Application No. 63 / 619,874, filed on January 11, 2024, and U.S. Provisional Patent Application No. 63 / 639,146, filed on April 26, 2024, the entire contents of which are each herein incorporated by reference for all purposes. STATEMENT REGARDING FEDERALLY SPONSORED RESEARCH OR DEVELOPMENT

[0002] Not applicable. TECHNICAL FIELD

[0003] This disclosure relates to the field of multidrug infusion pump control and management, and more particularly to systems and methods for delivery of complex multidrug infusions in, for example, anesthesia and critical care. BACKGROUND

[0004] Pump-driven intravenous infusion of life-critical medications is the standard of care in operating rooms and in many critical care settings. In situations where intravenous access is limited, as in the setting of a critically ill neonate, multiple drugs may be co-infused through a common fluid pathway. Multidrug co-infusion delivery systems typically consist of several drug delivery pumps along with a pump for dispensing a carrier fluid. During operation, the drugs and the carrier fluid enter a manifold, where they mix with each other at different points along a shared fluid flow path on their way to the delivery point. The delivery point being the tip of the intravenous catheter within a vein. The total volume of the fluid path between a drug’s point of entry into the shared fluid pathway and the delivery point is referred to as the dead volume (also termed common volume). In other words, the dead volume constitutes the space that the drugs must traverse before reaching the delivery point. The time 1 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) required for the drugs to traverse this dead volume can impart a considerable delay in achieving the targeted blood stream dosed set at the pump. This delay causes a mismatch between intended drug delivery profiles and the actual drug delivery profiles to a patient. This problem is clinically relevant at the start of an infusion, when dose changes are set at the pump, when drug containers are changed, and with the addition or cessation of a co-infused drug. SUMMARY

[0005] The present disclosure addresses these and other needs by providing systems, devices, methods, algorithms, and / or media for identifying piece-wise continuous mixed fluid flow regions within multi-drug infusion manifolds, dynamically assembling in memory the identified mixed fluid flow regions and solving the fluid drug transport therein with help of mathematical models subject to user specified operating parameters and constraints. The device can use this data for real-time visualizing the drug delivery profiles of a plurality of drugs, for dynamic guard railing in multidrug infusions, or for coordinated control of infusion pumps for better aligning the drug delivery profiles with clinical expectations. Relevant user specified parameters include structural details of infusion manifolds, the flow rate of a carrier fluid, the stock concentration and blood stream half-lives of infused drugs, and the intended drug infusion rates. Relevant constraints include hardware constraints such as the minimum and maximum pump infusion rates and fluidic constraints such as the permissible amount of excess drug and carrier fluid. A drug delivery profile, as used herein, refers to a rate of drug delivery at a delivery point over a period of time. Dynamic guard railing, as used herein, refers to dynamically calculated safety limits and associated warning mechanisms to avoid unintended drug over- or under-doses during multidrug infusions. Relevant clinical expectations include the immediate delivery of infused drugs to the patient at the intended dosage, minimal intermittency in drug delivery during container changes, and immediate cessation in delivery of stopped drugs.

[0006] According to one aspect of the present disclosure, a multidrug delivery system is provided. The system comprises a plurality of infusion pumps; an infusion manifold including a plurality of fluid input ports respectively connected to the plurality of infusion pumps and including a fluid mixing region; and a controller including at least one processor and a memory, the controller being operatively connected to the plurality of infusion pumps 2 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) and the infusion manifold and being configured to automatically perform operations comprising: modeling, in real-time, a spatiotemporal evolution of a drug concentration in the fluid mixing region in accordance with a transport model, and dynamically adjusting a rate of fluid flow from at least one of the plurality of infusion pumps to at least one of the plurality of fluid input ports of the infusion manifold based on the modeling.

[0007] According to another aspect of the present disclosure, a multidrug delivery management method is provided. The method comprises receiving a three-dimensional internal fluid volume of an infusion manifold, the infusion manifold including a plurality of fluid input ports respectively connected to a plurality of infusion pumps and including a fluid mixing region; modeling, in real-time, a spatiotemporal evolution of a drug concentration in the fluid mixing region in accordance with the transport model; and dynamically adjusting a rate of fluid flow from at least one of the plurality of infusion pumps to at least one of the plurality of fluid input ports of the infusion manifold based on the modeling.

[0008] According to another aspect of the present disclosure, a non-transitory computer-readable medium is provided. The non-transitory computer-readable medium stores instructions that, when executed by at least one processor of a multidrug delivery control device, cause the control device to perform operations comprising presenting a graphical user interface (GUI) to a user via a display of the control device; receiving, via the GUI, a first user input corresponding to an input operating parameter and a second user input corresponding to an input constraint; displaying, via the GUI, a visualization of a drug delivery profile based on the first user input and the second user input; and in response to a command to begin a multidrug delivery operation, instruct a multidrug delivery system to perform operations including: modeling, in real-time, a spatiotemporal evolution of a drug concentration in a fluid mixing region of an infusion manifold in accordance with a transport model, and dynamically adjust a rate of fluid flow from at least one of a plurality of infusion pumps to at least one of a plurality of fluid input ports of the infusion manifold based on the modeling. BRIEF DESCRIPTION OF THE DRAWINGS

[0009] Some examples of the disclosure are described herein with reference to the accompanying figures. The description, together with the figures, makes apparent to a person having ordinary skill in the art how some implementations of the disclosure may be practiced. 3 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) The figures are for the purpose of illustrative discussion and no attempt is made to show structural details of an example in more detail than is necessary for a fundamental understanding of the teachings of the disclosures. In the drawings:

[0010] FIG.1 illustrates an example fluid management card.

[0011] FIG. 2 illustrates an example relationship between drug delivery delay and physiological parameters.

[0012] FIG.3 illustrates an example the time scales associated with infusion mediated drug delivery.

[0013] FIG. 4 illustrates an example experimental setup in accordance with various aspects of the present disclosure.

[0014] FIG.5 illustrates example absorption spectra in accordance with various aspects of the present disclosure.

[0015] FIG.6 illustrates an example single drug concentration curve in accordance with various aspects of the present disclosure.

[0016] FIG. 7 illustrates an example validation of an algorithmic framework in accordance with various aspects of the present disclosure.

[0017] FIG. 8 illustrates an example single drug control in accordance with various aspects of the present disclosure.

[0018] FIG. 9 illustrates an example computational time effect in accordance with various aspects of the present disclosure.

[0019] FIG. 10 illustrates example infusion scenarios according to various aspects of the present disclosure.

[0020] FIG.11 illustrates an example comparison of various dispersion coefficients in accordance with various aspects of the present disclosure.

[0021] FIG. 12 illustrates example two drug control with a deterministic model in accordance with various aspects of the present disclosure.

[0022] FIG. 13 illustrates example reinforcement learning (RL) in accordance with various aspects of the present disclosure. 4 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)

[0023] FIG. 14 illustrates example validation rewards in accordance with various aspects of the present disclosure.

[0024] FIG. 15 illustrates an example single drug control with reinforcement learning in accordance with various aspects of the present disclosure.

[0025] FIG. 16 illustrates an example model architecture in accordance with various aspects of the present disclosure.

[0026] FIG. 17 illustrates example control trajectories in accordance with various aspects of the present disclosure.

[0027] FIG.18 illustrates an example pump in accordance with various aspects of the present disclosure.

[0028] FIG. 19 illustrates example device performance in accordance with various aspects of the present disclosure.

[0029] FIG. 20 illustrates an example system schematic in accordance with various aspects of the present disclosure.

[0030] FIG.21 illustrates an example impact of the infusion mechanism in accordance with various aspects of the present disclosure.

[0031] FIG.22 illustrates an example user interface in accordance with various aspects of the present disclosure.

[0032] FIG. 23 illustrates an example multidrug delivery management method in accordance with various aspects of the present disclosure. DETAILED DESCRIPTION

[0033] In the following detailed description, reference is made to the accompanying drawings in which specific examples are shown by way of illustration. These examples are described in sufficient detail to enable those of ordinary skill in the art to practice the disclosure. It should be understood, however, that the detailed description and the specific examples, while indicating examples of embodiments of the disclosure, are given by way of illustration only and not by way of limitation. From this disclosure, various substitutions, modifications, additions rearrangements, or combinations thereof within the scope of the disclosure may be made and will become apparent to those of ordinary skill in the art. 5 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)

[0034] Unless otherwise indicated, the various features illustrated in the drawings may not be drawn to scale. The illustrations presented herein are not necessarily intended to be actual views of any particular method, device, or system, but are merely idealized representations that are employed to describe various embodiments of the disclosure. Accordingly, the dimensions of the various features as illustrated may be arbitrarily expanded or reduced for clarity. In addition, some of the drawings may be simplified for clarity. Thus, the drawings may not depict all of the components of a given apparatus (e.g., device) or method. In addition, like reference numerals may be used to denote like features throughout the specification and figures.

[0035] It should be understood that any reference to an element herein using a designation such as “first,” “second,” and so forth does not limit the quantity or order of those elements, unless such limitation is explicitly stated. Rather, these designations may be used herein as a convenient method of distinguishing between two or more elements or instances of an element. Thus, a reference to first and second elements does not mean that only two elements may be employed there or that the first element must precede the second element in some manner. Also, unless stated otherwise a set of elements may comprise one or more elements.

[0036] Unless otherwise specified or indicated by context, the terms “a,” “an,” and “the” mean “one or more.” As used herein, unless otherwise limited or defined, “or” indicates a non-exclusive list of components or operations that can be present in any variety of combinations, rather than an exclusive list of components that can be present only as alternatives to each other. For example, a list of “A, B, or C” indicates options of: A; B; C; A and B; A and C; B and C; and A, B, and C. Correspondingly, the term “or” as used herein is intended to indicate exclusive alternatives only when preceded by terms of exclusivity, such as “only one of,” or “exactly one of.” For example, a list of “only one of A, B, or C” indicates options of: A, but not B and C; B, but not A and C; and C, but not A and B. In contrast, a list preceded by “one or more” (and variations thereon) and including “or” to separate listed elements indicates options of one or more of any or all of the listed elements. For example, the phrases “one or more of A, B, or C” and “at least one of A, B, or C” indicate options of: one or more A; one or more B; one or more C; one or more A and one or more B; one or more B and one or more C; one or more A and one or more C; and one or more A, one or more B, and one or more C. Similarly, a list preceded by “a plurality of” (and variations thereon) and 6 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) including “or” to separate listed elements indicates options of one or more of each of multiple of the listed elements. For example, the phrases “a plurality of A, B, or C” and “two or more of A, B, or C” indicate options of: one or more A and one or more B; one or more B and one or more C; one or more A and one or more C; and one or more A, one or more B, and one or more C.

[0037] As used herein, “about,” “approximately,” “substantially,” and “significantly” will be understood by persons of ordinary skill in the art and will vary to some extent on the context in which they are used. If there are uses of these terms which are not clear to persons of ordinary skill in the art given the context in which they are used, “about” and “approximately” will mean plus or minus ≤10% of the particular term and “substantially” and “significantly” will mean plus or minus >10% of the particular term.

[0038] As used herein, the terms “include” and “including” have the same meaning as the terms “comprise” and “comprising” in that these latter terms are “open” transitional terms that do not limit claims only to the recited elements succeeding these transitional terms. The term “consisting of,” while encompassed by the term “comprising,” should be interpreted as a “closed” transitional term that limits claims only to the recited elements succeeding this transitional term. The term “consisting essentially of,” while encompassed by the term “comprising,” should be interpreted as a “partially closed” transitional term which permits additional elements succeeding this transitional term, but only if those additional elements do not materially affect the basic and novel characteristics of the claim.

[0039] In some examples, the systems and methods set forth herein may be implemented on or using one or more computing devices, each of which includes a processor and a memory. As used herein, a “processor” may include one or more individual electronic processors, each of which may include one or more processing cores, and / or one or more programmable hardware elements. The processor may be or include any type of electronic processing device, including but not limited to central processing units (CPUs), graphics processing units (GPUs), application-specific integrated circuits (ASICs), field-programmable gate arrays (FPGAs), microcontrollers, digital signal processors (DSPs), or other devices capable of executing software instructions. When a device is referred to as “including a processor,” one or all of the individual electronic processors may be external to the device (e.g., to implement cloud or distributed computing). In implementations where a device has multiple 7 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) processors and / or multiple processing cores, individual operations described herein may be performed by any one or more of the microprocessors or processing cores, in series or parallel, in any combination. In some implementations, one or more of the processing units or processing cores may be remote (e.g., cloud-based).

[0040] As used herein, a “memory” may be any storage medium, including a non- volatile medium, e.g., a magnetic media or hard disk, optical storage, or flash memory; a volatile medium, such as system memory, e.g., random access memory (RAM) such as dynamic RAM (DRAM), synchronous dynamic RAM (SDRAM), static RAM (SRAM), extended data out (EDO) DRAM, extreme data rate dynamic (XDR) RAM, double data rate (DDR) SDRAM, etc.; on-chip memory; and / or an installation medium where appropriate, such as software media, e.g., a CD-ROM, or floppy disks, on which programs may be stored and / or data communications may be buffered. The term “memory” may also include other types of memory or combinations thereof. For the avoidance of doubt, cloud storage is contemplated in the definition of memory. A memory is an example of a non-transitory computer-readable medium which stores instructions that are executable by a processor (or processors), the execution of which causes the executing device (e.g., a computer) to perform certain operations, such as those operations described herein.

[0041] Reduction of medication errors has been a priority since the publication of U.S. Institute of Medicine’s “To Err Is Human” in 1999. Attempts have been made since then to mitigate such errors, including the development, licensing, and global implementation of infusion pumps with medication databases. As healthcare therapies become more sophisticated, including cardiac interventional procedures, organ transplantation, and survival of premature neonates, there exists a need to support front line caregivers delivering ever-more nuanced care.

[0042] The present disclosure addresses these and other challenges in dynamic care settings, particularly in operating rooms and intensive care units where multiple life-critical medications are administered by pump-driven intravenous infusion. Comparative examples involve the use of manually managed multidrug infusion systems comprising several drug delivery pumps and (usually) a pump for carrier fluid. The drugs and carrier fluid infused by the pumps mix and move in a hydrodynamically coupled fashion along a shared fluid flow path, comprised of a manifold and intravascular catheter, before being delivered into the bloodstream at the tip of an intravascular catheter. 8 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)

[0043] In typical comparative practice, clinicians manually adjust infusion pumps settings to achieve intended physiologic targets such as systemic blood pressure, cardiac output, heart rate, pulmonary artery pressure and blood flow to the brain and spine. Manually managing and programming life-critical infusions add a significant cognitive load and overhead to a clinician, particularly in dynamic clinical settings such as cardiac, vascular or brain surgery and the intensive care unit. In these clinically dynamic setting infusions of multiple life critical drugs may be started, stopped, or have dose adjustments in short time frames (e.g., 3 to 5 minutes or even less). At the same time, the total fluid delivered to the patient needs to be reduced to avoid fluid overload, particularly in a pediatric critical care or surgical setting. These dynamic high-stakes infusion management demands are further confounded by three factors: concentrated drugs flowing at 0.1-3.0 mL / hr to reduce fluid load, hydrodynamic coupling, and the “dead volume.” Hydrodynamic coupling refers to the phenomenon where the flow of drugs influence each other within the shared fluid flow path, mutually affecting their delivery kinetics. As a result, addition or removal of drug transiently affects the delivery of all the drugs in the system. The dead volume (sometimes identified as the “common volume”) refers to the total fluid volume from a drug’s entry point into the shared fluid pathway to the delivery point, representing the space the drugs must traverse to reach the patient. This dead volume (which can be up to 1.5 to 2 mL, or up to 10× the drug volume infused in an hour) can introduce a significant delay in achieving a desired dose-rate, leading to a mismatch between clinical intention and actual drug delivery profiles.

[0044] Manually managing multidrug infusions is a cognitively demanding task as multiple drugs may have to be started, stopped, or dose adjusted in short time frames, while tightly controlling the total fluid delivered to the patient. FIG.1 illustrates an example of a fluid management card for a hypothetical patient in a pediatric intensive care unit, and demonstrates this complexity.

[0045] Lacking specific knowledge of real-time drug levels within the dead volume, and consequently the rates of entry of a mass of drug into the blood stream, clinicians utilizing the comparative examples are forced to manually adjust drug flow rates on multiple infusion pumps by serial, intuitive, approximations and manual titration to overcome drug delivery delays to achieve a desired clinical effect. This can lead to suboptimal outcomes such as severe oscillations in vital physiologic parameters including, as stated above, systemic blood pressure, 9 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) cardiac contractility, or blood flow to the brain. Indeed, the training requirements in these “artisanal nuances” of pump adjustment requires a lengthy process of learning by new clinicians before achieving “mastery.” Even experienced and expert clinicians may be unable to achieve precise rates of drug delivery in the intended time frame (just a few minutes) targeted to specific clinical endpoints in dynamic clinical settings. Ultimately, manually managed multidrug infusions utilizing medical grade pumps, particularly at low infusion rates, can be a major source of physiologic instability, undesired outcome and / or medication errors resulting from unpredictable delivery profiles and delivery delays.

[0046] FIG.2 illustrates an example of the impact of drug delivery delay on controlling important physiological parameters, and shows both drug dose per unit time and the change in blood pressure as a function of time. In FIG. 2 and throughout this disclosure, drug infusion begins at t = 0 unless otherwise indicated in the plot. FIG.2 shows a mathematical correlationbetween the change in pressure Δ^^ and the drug concentration ^^ௗ given by the formula Δ^^ ൌ^^^ బ.యషర. There are two important time scales that dictate how quickly a drug acts. A drug ^ା൬^^^ as a bolus takes time depending on its pharmacokinetic (PK) characteristics topartition into the tissues of interest, and act on appropriate cellular targets to elicit the appropriate biochemical response. FIG. 3 illustrates the relationship between infusion pump control and the relative magnitude of the absorption time scale ^^^(corresponding to the time it takes for a drug to act on cellular targets and elicit the appropriate biochemical response after it enters the body) and the Poiseuille time scale ^^^(due to advective transport of the drug within infusion manifolds). FIG.3, image (a) is a schematic showing the PK of an administered drug, which dictates ^^^. FIG. 3, image (b) shows the variables (Qc, Qd, Vdead) that dictate ^^^during catheter mediated deliver of drugs where Qc is carrier flow rate, Qd is the flow rate of drug entering the common fluid pathway and Vdeadis the dead volume as defined above). FIG. 3, image (c) shows the normalized concentration (concentration ^^ as a fraction of the steady state concentration ^^^) evolution as predicted by the Poiseuille time scale. The Poiseuille time scale enters the problem purely due to the advective transport of the drug within the catheter. For acatheter with a dead volume ^^ௗ^^ௗ and total drug flow rate ^^௧^௧ ൌ ^^^ ^ ^^ௗ, the time scale forthe drug to reach the catheter tip scales as ^^^ ൌ ^^ௗ^^ௗ / 2^^௧^௧ and the concentration scales as10 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)^^ / ^^^ ൌ 0 if ^^ ^ ^^^ and ^^ / ^^^ ൌ 1 െ ^^^ / ^^ if ^^ ^ ^^_^^. The factor of 2 comes from the fact thatmaximum velocity in Poiseuille flow is twice the mean velocity.

[0047] For many drugs, the absorption time scale ^^^is known, and among other things depends on the psychochemical characteristics and the pharmacokinetics of the drug. A sample of known drug time scales is shown in Table 1. Table 1 comprises data for a carrier flow rate of 10 mL / hr (adult) or 1.5 mL / hr (pediatric), a dead volume of 1.2 mL (adult) or 0.7 mL (pediatric), or a patient weight of 70 kg (adult) or 7 kg (pediatric). Table 1 – Therapeutic Action Timescale and Deliver Information of Selected Drugs # Name ^^^(min) Stock conc. (mcg / mL) Delivery rate11 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)

[0048] Controlling infusion pumps becomes important when ^^^ / ^^^ ^ 1. Herein, thequantity ^^^ / ^^^is referred to as the Shafer number ^^ℎ. This quantity may also be referred to as the Forssmann number ^^^^. There are at least two scenarios where this can happen in the clinic: (a) for fast acting drugs (low ^^^) such as anesthetics and vasopressors, and (b) for infusions atsmall flow rates (low ^^௧^௧) such as in neonatal care. At a low Shafer number (≪ 1), drugtransport kinetics through the catheter effectively dictate the time taken for an infused drug to display its therapeutic effect. These conditions benefit from active management of drug infusions. From the definition of the Shafer number, an additional scenario can be identified: infusions through geometries with large dead volume (high ^^ௗ^^ௗ). Many stopcock manifold assemblies common in both adult and pediatric intensive care settings have large dead volumes. Taken together, these three scenarios account for a significant fraction of all anesthetizing and intensive care situations.

[0049] The present disclosure tackles the above and other problems by developing deterministic and machine learning based algorithms for dynamically assembling fluidic paths in memory, computing spatiotemporally resolved drug concentrations within the assembled fluid paths, and leveraging this information for controlling intravenous drug delivery to satisfy clinical objectives. This framework, referred to herein as SMART (Synchronized-pump Management Algorithms for Reliable Therapies), presents a solution to pervasive problems in multidrug delivery; for example, interconnected infusion devices synchronously managed by a central processor interfaced with an intuitive user interface. The disclosed system can provide unprecedented graphical and numeric data to the clinicians including real-time drug concentrations within infusion manifolds and dead volume, and the real-time rates at which the drug mass is entering the blood stream. This framework can also control the connected infusion devices to execute clinical objectives such as minimizing the drug delivery delay induced by the dead volume. For example, in a simulation of pediatric critical care, where every second counts, the framework set forth herein reduced the dead volume delay in drugs entering the body, from over 30 minutes to less than 5 minutes. SMART not only reduces drug delivery delays but also provides improved performance for synchronized drug adjustments, drug container changes, and during the addition or cessation of co-infused drugs. The intelligent algorithm set forth herein anticipates the impact of these changes on the drug delivery profile, making preemptive adjustments to other hydrodynamically coupled co-infused drugs to 12 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) maintain a steady delivery of all actively infused drugs. This ensures a continuous and reliable delivery of medications, directly aligning with the clinical expectations and enhancing patient safety. This framework also enables the continuous documentation of temporally resolved drug delivery rates, both executed on the infusion devices and at the delivery site. Finally, the tight control of infusions permitted by SMART makes it possible to employ more-concentrated drug formulations, which can significantly improve the problem of patient fluid overload, the portability of pumps and medications and their administration in austere environments. It is to be noted that, while the following description focuses primarily on implementations of the system with liquid media for purposes of explanation, the present disclosure is not so limited. In some implementations, the SMART methodology and associated systems set forth herein may operate with gases, liquids, granular media, or any combination thereof.

[0050] The SMART framework may be implemented in a variety of devices. One implementation is a device that can read an input from a slider or related graphical input and determine a tradeoff between the delivered excess carrier fluid volume and the reduction in the dead volume transport delay. As outlined above, as ^^ ^^^ೌ^^ൌ ଶொ^^^. Thus, the lag time to achieve stead state delivery is a function of total system fluid flow. A high total system fluid flow, for example from a high carrier fluid flow rate, results in a smaller lag time, at the expense of increasing the delivery of fluid to the patient. The SMART system provides the clinician control over this trade-off. In one example, the slider is set to 0, implying there is no tradeoff. The maximum reduction in dead volume delay is obtained with the least possible excess carrier fluid that guarantees performance. In another example, the slider is set to 0.5 for a patient that cannot handle excess fluid. A moderate reduction in dead volume delay is obtained, albeit with a smaller excess carrier fluid volume.

[0051] Another implementation is a device that can computationally identify piecewise sections of mixed fluid flow, dynamically assemble sequentially in memory the identified mixed flow regions, and sequentially solve the spatiotemporal evolution of the drug(s) along the mixing sections with the help of a transport model. In one example, a single drug and carrier fluid is infused into a four-stopcock manifold assembly and then through a central venous catheter to the patient’s circulation. The device dynamically assembles by computation the piecewise mixing section containing the drug and the carrier fluid in memory. The spatiotemporal evolution of the drug within the mixing section is solved with the help of a 13 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) transport model. In another example, the same situation exists but a clinician introduces a plurality of drugs into the existing four-stopcock assembly. As used herein, a “stopcock assembly” refers to a plurality of stopcocks connected in series to form a manifold. The device dynamically calculates and updates the sequential piece-wise mixed flow sections introduced by the new drug. The device calculates the spatiotemporal evolution of the new drugs accounting for the previously existing spatiotemporal distributions of drug concentrations.

[0052] Yet another implementation is a device that can perform drug sequence independent constrained non-linear optimization for reducing dead volume transport delay based on user specified target objectives. In one example, the device performs non-linear optimization with the objective of reducing the difference between the actual and set drug delivery rates. In another example, the device performs non-linear optimization with the objective of reducing the difference between the actual and set area under the curve (AUC).

[0053] Experimental Analysis

[0054] In developing and testing SMART, a high-fidelity framework was developed, comprising of an experimental platform and OpenFOAM based fully resolved numerical simulation for accelerated development and testing of SMART. FIG. 4 illustrates an example of the scenario and setup. In FIG.4, image (a) is a schematic of the clinical scenario, and shows drug infusion pumps (in this example, syringe pumps) connected to a fluid line that is in turn connected to a patient, and a carrier (e.g., saline) solution pump connected to the fluid line via a stopcock assembly. While the illustrated example corresponds to syringe pumps, in various implementations of the present disclosure any type of infusion pump, or any combination of multiple types of infusion pumps, may be used. FIG.4, graph (b) shows the Shafer number of selected drugs administered in anesthesia and critical care. For calculating the Shafer number, the lower bound of the delivery rate (see Table 1) was used. Drugs administered in under “adult” concentrations are shown using darker labels than drugs administered under “pediatric” conditions. Most of the tightly titrated drugs used with the SMART methodology set forthherein have low Shafer numbers (^^ℎ ^ 1). At low Shafer numbers, drug transport kineticsthrough the catheter effectively dictate the time taken for an infused drug to display its therapeutic effect. It is typically these conditions that are most prone to delivery delays, and that consequently benefit from active management of drug infusions in the clinic. Thus, the 14 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) Shafer number presents a physically grounded metric to identify scenarios that most benefit from drug delivery management by SMART, without relying on clinical experience.

[0055] FIG.4, image (c) shows the benchtop experimental setup, and includes pumps connected to an output, and a light source and a spectrophotometer connected to the output via a Z-flow cell and a stopcock assembly. FIG.4, image (c) shows the computational framework used to validate the developed device and algorithms. Thus, the experimental platform mimics a bedside central venous catheter setup. The tip of the catheter is interfaced with a low volume Z-flow cell to provide a large optical path for enhancing SNR of spectroscopic measurement while minimizing additional dead volume in the system. Food grade dyes Erioglaucine (Sigma Aldrich, 861146-25G) and Tartrazine (Sigma Aldrich, T0388-100G), which have distinct spectral absorption signatures, are used as model dyes. FIG.5 illustrates the spectral absorption signatures. In FIG. 5, graph (a) presents the absorption spectrum of Methylene Blue and Tartrazine, and graphs (b) and (c) showing the concentration versus normalized transmission intensity calibration curves for Methylene Blue and Tartrazine, respectively. Solid curves correspond to single exponential fits while dashed lines correspond to linear fits.

[0056] The benchtop spectrophotometric setup used for measuring drug concentrations at the catheter tip consists of two high intensity light sources capable of emitting light at wavelengths of 630 nm (I630) and 430 nm (I430) (Thorlabs M625l4 and M430l4, respectively). Both LEDs are integrated into the system using a Thorlabs C4W 30 mm Cage Cube, which includes a dichroic mirror (Thorlabs DMLP505R) and a specialized lens (Thorlabs LA4647). This assembly is able to combine and converge the light beams from both LEDs and couple them to the inlet optical fiber cable that is connected to a Z flow cell (Ocean Insight FIA- ZSMA-SS). The LEDs are powered and controlled by a Thorlabs LEDD1B T-Cube LED Driver, allowing for precise modulation of light intensity. The emitted light enters the flow cell through an optical fiber, passes through the fluid in the optical path before being transmitted to the spectrometer via another optical fiber. The employed Thorlabs CCS100 Compact Spectrometer was used to record intensities in the wavelength range spanning 350 to 700 nm with an integration time of 1000 μs. The recorded spectrophotometric data was processed as detailed in FIG. 5 to recover the temporally resolved drug concentration at the catheter tip. A digital scale (Ainsworth M-220) was used to measure the total dispensed fluid mass as function of time. This data was further processed to obtain the total infused flow rates. 15 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)

[0057] In FIG.5, graphs (b) and (c), the solid dots represent measurements at 424 nm and the open circles represent measurements at 631 nm. As the concentration of dye increases, the normalized light transmission decreases due to increased absorbance by the dye. This is consistent with the Beer-Lambert law, which states that the absorbance of a substance is directly proportional to its concentration. By fitting these data points, the following expression can be derived to describe how the intensity of light at each wavelength changes with dye concentration. ^^ ି^.^^^ଶ^^ଷ^ா ൌ ^^ ಶ^^ସଷ^ா ൌ െ0.532^^ா ^ 1^^ସଷ^் ൌ ^^ି^.^^ଽ^^^^^^ଷ^் ൌ െ0.00001339^^் ^ 1

[0058] In the above expressions, the subscript E refers to Erioglaucine and the subscript T refers to Tartrazine. To determine the concentrations ^^்and ^^ா, one must solve a system of coupled equations based on the experimentally measured resultant intensities (^^^ଷ^and ^^ସଶସ) at the two respective wavelengths such that the following is satisfied. ^^^^ଷ^ ൌ ^^^ଷ^் െ ^^^ଷ^ா

[0059] Returning to FIG.experiment a computation platform running SMART is connected to infusion pumps loaded with dye(s), and performance is evaluated by measuring temporal evolution of the drug concentration at the catheter tip. For the experimental analysis, LEGATO 110 (KD Scientific) syringe pumps were used. The 4- channel peristaltic pumps were constructed utilizing linear peristaltic units extracted from Baxter 6200 pumps. For accelerating development and orthogonal validation of SMART, a fully resolved numerical simulation was performed using OpenFOAM. A catheter-stopcock manifold assembly identical to that used in the experiments (see row (i) of FIG.4, image (c)) was microCT scanned (X-Tek HMXST225) (see row (ii) of FIG. 4, image (c)) and processed in ORS DragonFly (Comet Technologies) to extract the internal fluid volume (see row (iii) of FIG. 4, image (c)). The extracted 3D volume was saved as an STL file and appropriately meshed using the snappy hex mesh, and solved using a custom advection-diffusion equation 16 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) solver based on native IcoFoam OpenFOAM to include a generalized advection-diffusion equation for solving the transport of a passive scalar with adaptive time stepping. The simulations were parallelized by decomposing the meshed geometry using decomposePar (OpenFOAM) and run across 30 compute cores. A mesh independence study was performed prior to using the model for evaluating SMART.

[0060] In particular, to evaluate the dependence of mesh characteristics on the numerical results, four different meshes were created with the number of cells varying an order of magnitude from 50K to 550K. Identical single drug test cases were run on all the meshes with carrier and drug stream flow rates of 10; mL / hr and 3; mL / hr respectively. Table 1 shows this in more detail. The evolution of single drug concentration at the tip of the catheter was measured to evaluate mesh independence. The negligible difference in concentration profiles between meshes having ~350K and ~550K elements indicates that numerical mesh is achieved at a size of ~350K. This mesh was used for all subsequent calculations described herein. Table 2 – Mesh Independence Study Metrics No. Cells Single Core Equivalent Clock Time (s) Cd (t = 500 s) tcd90

[0061] FIG. 6 illustrates the evolution of single drug concentration at the tip of the catheter for the four different meshes with varying degrees of refinement. The carrier stream was injected at 10 mL / hr while the drug was injected at 3 mL / hr. The results show that mesh independence is achieved at a size of ~350K.

[0062] Deploying SMART for real-time control of drug infusions in clinical settings is facilitated by computationally efficient algorithms that can be executed in small footprint hardware. Accomplishing this may be facilitated by developing accurate reduced order models for accurately tracking drug transport within the catheter. Various mathematical formulations may be used for tracking drug concentrations. The advective transport of drugs within catheters 17 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) may be described by the convection-diffusion equation, for example as set forth in the following Equation 1. ^^^^^ ^ ^^.∇^^ ൌ ^ ଶ (1)^^^^^ ^^∇ ^^^

[0063] In Equation 1, thei-th drug in the catheter. The velocity field ^^ satisfies the Navier Stokes equation, as set forth in the following Equation 2. ^^^^ 1 ^^ ଶ^^^^^.∇^^^ ൌ െ^^∇p ^ ^^∇ ^^ (2)

[0064] General analytical solutions are tractable for the above equations for unidirectional velocity fields, and one can solve the equations accurately utilizing numerical packages such as OpenFOAM. These simulations, referred to as fully resolved simulations, are computationally expensive. For improving computational performance without compromising accuracy, examples of the present disclosure employ reduced order models derived from the above equations. The generalized 3D advection diffusion equation (ADE) that governs drug transport can be computationally simplified by neglecting radial and azimuthal variations and focusing solely on the axial evolution of the drug concentration. Such an approximation may be realized by utilizing the Taylor diffusion approximation, which holds for fully developed Poiseuille flows, constant and isotropic species diffusivity, and for geometries where length^^ ൌ 0.2^^^^^^. Variations in the radial extent of the stopcock-catheter assembly is capturedthrough a network model with piecewise continuous cylindrical units with matched outlet-inletboundary conditions. In one example for a Poiseuille flow field in cylindrical coordinates ^^ ൌ^^^^^^, the Taylor diffusion approximation is deployed as set forth in the following Equation 3. ^^^ഥ^ప ^^^ഥ^^^^^ଶ^^ଶ^ഥ^^ ^ ప ^ ^^ ^^ ^ ^ 1^^ ప^ (3)

[0065] In,^ ,Diffusion coefficient (or ^ഥ^ప is the radial-average initial the developed algorithmic framework on canonical geometries. In FIG. 7, graph (a) shows spatio-temporal drug concentration profiles obtained from fully resolved simulations using OpenFOAM; graph 18 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) (b) shows normalized drug concentration evolution at different points along the length of the cylindrical geometry; and graph (c) shows normalized drug concentration evolution at the end of the geometry where the drug is introduced orthogonal to the carrier stream. Graph (c) also shows the effect of SMART in improving the drug delivery kinetics in the cylindrical geometry.

[0066] The reduced order model is solved numerically using the second order accurate Crank-Nicholson finite difference scheme. Numerical oscillations are minimized by employing the Backward-Euler scheme to evolve the first step, where sharp gradients in drug concentration are present. Sparse matrix implementation of the schemes is adopted to significantly enhance the computational and memory management efficiency. Overall, the adopted reduced order model and the associated solution methodology provide a 104enhancement in computational time per physical core as compared to fully resolved numerical simulation, while providing accurate resolution of the drug delivery profiles across a range of infusion conditions prevalent in anesthesia and critical care. These profiles also reveal that the extent of delivery delays scale with the Shafer number, with situations common in pediatric settings requiring over 30 minutes to reach the desired delivery rates at the catheter tip. FIG.8 illustrates the single drug control, FIG. 9 illustrates the improvement in computational time, and FIG.10 shows the profiles across selected infusion conditions.

[0067] In FIG. 8, graph (a) shows the baseline drug infusion kinetics for three cases without SMART. The indicated Shafer numbers are calculated for a drug behaving likepropofol (^^^ ൌ 40 s). FIG.8, graph (b) is a heatmap that shows the discrete state-action values^^^^^ప , ^^^ (see Equation 1) evaluated at preferred values of the action space ^^^. The solid lineshows the effective policy ^^ௗ^^^^^^ప ^ executed by the deterministic framework. FIG.8, graph (c)presents a demonstration of improvement in drug delivery with the developed algorithms forthree different scenarios explored in graph (a). For all cases, ^^^^^௫ / ^^^ ൌ 3. FIG. 8, graph (d)illustrates phase space showing the improvement in time to reach 90% of set drug concentration (t90) as a function of variables in the control algorithm ^^௧and ^^^^^௫ / ^^^. The quantity ^^௧isa parameter ranging from 0 to 1 that controls the amount of excess fluid, with ^^௧ ൌ 1corresponding to no excess fluid. No improvement in delivery kinematics is physically possiblewhen ^^^^^௫ / ^^^ ൌ 0 or when ^^௧ ൌ 1. FIG. 8, graph (e) is an experimental and numericalvalidation of data shown in graph (d). FIG. 8 graph (f) illustrates phase space showing the 19 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) excess fluid as a function of variables in the control algorithm ^^௧and ொ^^ೌ^ொ^. FIG.8, graph (g) is an experimental and numerical validation of data shown in graph (f).

[0068] In FIG.9, the computational time required for simulating 1000 seconds of drug infusion for different computational models is shown. FIG.9, image (a) is for a fully resolved model solved using OpenFOAM; FIG.9, image (b) is for a 1-D reduced order model employing the Taylor diffusion approximation solved using MATLAB; and FIG. 9, image (c) is for a computationally optimized reduced order model solved using MATLAB. It can be seen that reduced order modeling improves computational time by a factor exceeding 104.

[0069] The selected profiles illustrated in FIG.10 also reveal that the extent of delivery delays scale with the ^^ℎ number, with situations common in pediatric settings requiring over 30 minutes to reach the desired delivery rates at the catheter tip for manual startup from an initial flow rate of zero. The SMART framework set forth herein reduces drug delivery delays via, for example, an ensemble of deterministic and reinforcement learning based approaches.

[0070] For non-standard situation where the geometry length falls below the aboveexpression (i.e., ^^ ^ 0.2^^^^^^), the Gill-Sankarasubramanian approximation may be applied tothe ADE for improved accuracy. Unlike the Taylor diffusion approximation, the Gill- Sankarasubramanian approximation is accurate at arbitrarily small time scales, as illustrated in FIG. 11, which shows that the temporally varying diffusion coefficient obtained via the Gill- Sankarasubramanian approximation smoothly evolves from the intrinsic diffusion coefficient of the liquid to the Taylor dispersion coefficient over time. In such non-standard situations, the accuracy and performance may further be improved by expressing the effective diffusion coefficient as a function of time as shown in the following Equation 4. ଶ ^ ^^^^ ^^^ ^^^^^ ^^ ^^ ^^ఛ^^ (4)

[0071] In^^^^ଷ^^^^^^^ଶ^^^^^^20 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) ^^^^^ ^^^ൌ ^^ଶ

[0072] Additionally, various formulations may be used for controldecisions. At its core, the developed policy chooses actions that minimizes the error between the expected and actual drug delivery rates at a given instant of time. The possible actions are independently varying the drug pump infusion rate ^^ௗ,^and the carrierinfusion rate ^^^ for each state defined by the drug concentration ^^^^^ప at the catheter tip. Thereward ^^൫^^^^^ప , ^^^ ,^^ௗ,^൯ may be simply defined as the negative of error in drug delivery. Onepossible example of the constrained multi-objective optimization reward function is described in Equation 5 below. ^^൫^^^^^ప , ^^^ ,^^ௗ,^൯(5)

[0073] Above, ^^^^^ప is the non-dimensional concentration of the i-th drug, ^^^ and ^^^ arerespectively the instantaneous and the set flow rates of the carrier fluid, ^^ௗ,^and ^^ௗ,^are respectively the instantaneous and the set flow rates of the carrier i-th drug, ^^௧is a non- dimensional weighting parameter that is bounded between 0 and 1 and constrains the excess carrier fluid administered to the patient, ^^ is a parameter that determines the permissible overshoot of the i-th drug, and ^^ is a parameter that determines the fluctuation in concentration of the i-th drug within he shared fluid path. Equation 5 is subject to the following conditions. 0^ ^^^ ^ ^^^^^௫0 ^ ^^ௗ,^ ^ ^^ௗ^^௫,^^^ௗ,^ ^ ^1 ^^^ௗ,^^^ௗ,^ ^ ^^ ^ ^^^ ^^ௗ,^ ^ ^^^^^ௗ,^ ^ ^1 ^^^ௗ,^^^ െ ^^ௗ,^ ^ ^^^ ^^ௗ,^ ^ ^^^21 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) ^^ௗ,^^^^^^ప ^ ^^^^ௗ,^

[0074] The first two the range of syringe flow rates that aredeemed clinically safe (or due , the next two constraints limit fluctuation in concentration at the inlet relative to the steady state value, and the last constraint places an upper bound on the permissible overshoot in drug concentration at the outlet. For improved stability during nonlinear optimization and for machine learning based algorithm development, in practice a non-dimensional version of Equation 1 is solved where the action space is bounded between -1 and 1. To effect this, the action variables may be normalized as shown in Equations 6a and 6b below. ^^ െ ^ì^ ^^െ ^^ , if ^^^^^௫ ^ ^^^ ^ ^^^^

[0075] and 1, with preferred actions tending to 0 as the drug concentration approaches steady state at the outlet. Equation 5 in the normalized action space can be written as the following Equation 7. ^^൫^^^^^ప , ^^^, ^^ௗ,^൯ൌ െ^^^ି^^ௗഢ ି^^,^^^ௗഢ ொௗೞ^ೌ^^,^ െ ൫1 ^ ^^^^^ଶ ప൯^^௧^^^^ଶ(7)

[0076] Equation 7െ1 ^ ^^^^ ^ 1െ1 ^ ^^^ௗ,^ ^ 1^^^ௗ,^^^^^^^^^^,^^^ு,^ െ ^^^^^^^^^^^^^ ^ 1 െ ^^ு,^^^^ௗ,^^^^^^^^^^,^^^^,^ െ ^^^^^^^^^^^^^ ^ ^^^,^ െ 122 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) ^^^ௗ,^^^^^^^^^^,^ ^ ^^^ െ ^^^^^ప

[0077] Above, the^^^^^௫ െ ^^^^^^^ , if ^^^^^௫ ^ ^^^ ^ ^^^^^^^^ ൌ ^^^^ௗ,^

[0078] State-Action valuescorresponding to the non-dimensional ^^^^^ప , ^^^ௗ,^ state-action space. This graph shows a slice ofthe 3D state-action value matrix evaluated at preferred values of the ^^^^action space. The hashed regions show regions that are off limits due to the imposed constraints. The deterministic policy ^^^ௗ^ at a given state returns actions with the largest state-action value; that is, argmax^^൫^^^^^ప , ^^^ ,^^ௗ,^൯. This function is shown by the solid line in FIG. 8, graph (b). In^ practice, to avoid computing the entire state-action value matrix, the preferred actions ^^^,^^ௗ,^given by the deterministic policy are recovered via sequential quadratic programming using publicly available libraries.

[0079] The deterministic control algorithm accelerates drug delivery to the patient, as demonstrated by the dramatic shift in drug delivery profiles from blue curves (w / o control) to the corresponding green curves as shown in FIG. 8, graph (c). Analytical delivery profiles expected from the deterministic control algorithm agree well with those obtained from experiments and fully resolved numerical simulations, demonstrating the accuracy and robustness of the developed framework. As implied in the equations above and demonstrated in FIG.8, graph (d), the enhancement in drug delivery depends both on parameter ^^௧and the relative magnitude of the maximum and steady state saline (^^^^^௫\^^^) infusion rates. The 23 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)degree of control depends linearly on ^^௧, with no control at ^^௧ ൌ 0 and maximum control at^^௧ ൌ 1. ^^^^^௫\^^^ determines the maximum enhancement possible. For instance, at ^^௧ ൌ 1and ^^^^^௫\^^^ ൌ 5, the time to attain 90% of steady state drug concentration (t90) is 10 timessmaller compared to the case without control. These analytical expectations are corroborated by experimental and fully resolved numerical data, as shown in FIG. 8, graph (e) for the caseof ^^^^^௫\^^^ ൌ 3.

[0080] The improvement in drug delivery time may result in a cost of delivering excess fluid to the patient, as shown in FIG.8, graph (f). However, because the control algorithms are able to achieve steady state concentrations by operating for a relatively short duration, theamount of excess fluid delivered is small. For instance, at ^^௧ ൌ 1 and ^^^^^௫\^^^ ൌ 5, a 10×improvement in drug delivery is accompanied by a cumulative excess fluid volume of about 0.75 mL, which makes such systems and methods beneficial for accelerating drug infusions in both adult and pediatric patients. This framework is scalable to multiple drugs infused in any arbitrary combination of ports as illustrated in FIG. 12. In FIG. 12, graph (a) illustrates two drug infusion without control, and graph (b) illustrates two drug infusion with control using the deterministic model.

[0081] While the “greedy” deterministic approach gives satisfactory performance for managing drug infusions within manifolds, there are opportunities to further improve drug infusion kinetics with non-greedy globally optimal approaches, particularly for cases with two or more drugs. Moreover, advances in personalized medicine are increasingly demanding scalable tools that can reliably manage drug infusions end-to-end to elicit a desired physiological response. Engineering deterministic models to attain these complex objectives is not a scalable approach. Motivated by these needs, the present disclosure further present a scalable reinforcement learning framework for catheter mediated drug delivery utilizing a dual buffer strategy: one persistent buffer with expert trajectories obtained from the deterministic model and a first-in-first-out experience buffer, as illustrated in FIG.13.

[0082] FIG. 13, graph (a) shows scalable reinforcement learning architecture for end- to-end control of infusion pump-based drug delivery utilizing a dual buffer strategy, with one persistent buffer with expert trajectories obtained from the deterministic model and a first-in- first-out experience buffer. The developed “Drug Infusion Gym” follows the OpenAI gym standard. FIG. 13, graph (b) is a heat map showing the State-Action-Values learned by RL 24 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)agent in the ^^^ௗ െ ^^^ௗ plane for single drug delivery (^^ௗ^^௫ / ^^ௗ ൌ 3). The solid line shows thelearned policy ^^ோ^. Graph (c) of FIG.13shows that the learned policies generate distinct action sequences with minimal inputs to inactive ports, as confirmed from a t-SNE embedding of action sequences executed by a trained agent for a single drug injected across different ports(^^ௗ^^௫ / ^^ௗ ൌ 3). In graph (d), the normalized time difference in attaining 90% of setconcentration between the reinforcement learning agent (t90ML) and the deterministic framework (t90DET) for single drug delivery as a function of maximum flow rate factor (^^ௗ^^௫ / ^^ௗ) and excess fluid weight (^^௧) is shown. Blue regions indicate conditions where the reinforcement learning agent outperforms the deterministic framework. FIG. 13, graph (e) shows he normalized time difference in attaining t90 for all infused drugs when simultaneously administering two drugs. Graph (f) is a comparison of multidrug delivery kinetics by SMART running either the deterministic framework or the reinforcement learning framework.

[0083] The above strategy not only helps in balancing forgetting and generalization, but also enhances the rate of learning and the ultimate performance of the RL agent as illustrated in FIG. 14. As can be seen in FIG. 14, expert training improves both the rate of learning and the ultimate performance of the RL agent.

[0084] For training the RL agent, a drug infusion gym was constructed following the OpenAI gym standard. The drug infusion gym employs the above-described reduced order model drug transport model for physically grounded and computationally efficient state evolution of the system in response to an agent’s interactions. The infusion gym includes a 24×1 observation space and a 4×1 action space. The agent’s observation space comprises six state variables, including a binary representation of the ports with drugs, the non-dimensionalposition within the manifold where the drug concentration is at 5% of the set value (^^^ହ^ ), thevelocity of ^^^ହ^ , the non-dimensional drug concentration at the catheter tip (^^^ௗ^1,non-dimensional drug flow rate, and the radio ^^ௗ^^௫,^ / ^^ௗ,^. The methods include step(action): for evolving the environment through one time step, reset(): for resetting the environment to the initial state, and render(): for visualizing the actions of the agent. Additional functionalities include routines to evoke the deterministic expert trainer and run a trained agent. In response to an action, the agent receives a reward comprised of four parts, namely: a penalty for large changes in action, a penalty for overshooting the drug delivery rate above a threshold, a reward for moving the drug closer to the patient, and a reward for reaching the desired delivery rate. 25 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)

[0085] The action penalty is a two-component negative reward that penalizes large changes in actions as well as actions on ports without drugs. The overshoot penalty penalizes the agent when the drug delivery overshoots beyond 50% of the set delivery rate. The forward reward rewards the agent when the drug delivery rate is made to approach the set rate. The destination reward rewards the agent when both the drug flow rates and pump infusion rates reach the set value. Example implementations of the four parts are set forth as the pseudocode below. actionPenalty = -np.sum(abs(self.prevAction-action * penaltyFactor) / self.numPorts – offTargetPenaltyFactor * np.sum(abs((1 –

[0086] In addition to graph (a) of FIG.13, this is further illustrated in FIG. 15. In one example implementation, a Twin Delayed Deep Deterministic Policy Gradient (TD3) network is employed as the reinforcement learning agent. The TD3 agent samples the dual buffer with a 20:80 (expert:experience) split as it learns a policy ^^ோ^that maximizes the Bellman equation from a given state ^^^, which is represented in the following Equation 8. ^^ோ^^^^^^ ൌ argmax^^^௧^^^^^^ା௧ା^,^^^ା௧ା^^ ^(8)௧

[0087] In Equation 8, ^^ is the discount factor, ^^ is the time relative to the given state ^^^, and ^^ is the agent’s action that specifies the non-dimensional drug flow rate. The continuous action space is bounded between [1, -1], with 1, 0, and -1 respectively corresponding to the maximum, set, and minimum allowed drug flow rates.

[0088] The presented drug infusion gym allows a TD3 agent to rapidly learn effective policies that respect desired constraints. This can be seen in FIG.13, graph (b), where a part ofthe state-action-value space is visualized for a single drug infusion in the ^^^ௗ െ ^^^ௗ plane. Thehatched area corresponds to regions that violate the constraints specified above. The solid line shows the action selected by the learned policy ^^ோ^^^^^^. To ensure safety redundancy, during 26 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) deployment, the actions of the model are guard-railed through a simple deterministic wrapper that enforces the constraints. The action space is constrained between [-1, 1], where the flow rates are non-dimensionalized according to expressions 6a and 6b above. As a result, the model is hardwired to respect the constraints on expression 5 above. The remaining constraints are softwired in the model through the rewards specified above. To ensure that the model does not violate these constraints, a simple wrapper is provided that limits the model actions to stay within the system constraints.

[0089] This is illustrated in FIG. 16, which shows the model architecture. The reinforcement agent has an actor-critic architecture with each network having 72960 neurons. The agent has a split buffer, which is partly populated by expert trajectories obtained from the deterministic model described above with a total batch size of 2560. The learned policies also generate distinct action sequences with minimal inputs to inactive ports. This is evident in FIG. 13, graph (c), where action sequences generated for a single active drug injected in different ports start off in distinct locations within a tSNE embedding, before converging to a common point at long times when the drug delivery has reached the set value (converging to an action of 0 in all ports).

[0090] The trained TD3 agents yielded comparable performance with the deterministic model for single active drug injection, as can be seen in graph (d) of FIG. 13 and in FIG. 15. However, the TD3 agent was dramatically superior to the deterministic model for the more complex case of multidrug delivery, as can be seen in graphs (e) and (f) of FIG.13 for the case of two drugs. This is a result of the trained TD3 agent’s ability to execute non-greedy decisions (see FIG.17, showing control trajectories executed by the RL for two drug infusions) and drive drug infusion trajectories that globally minimize the total drug delivery delay.

[0091] Example Implementations

[0092] The above experimental analysis illustrates several of the benefits in implementing the SMART modality to manage multidrug delivery using infusion pumps. FIG. 18 illustrates an example realization of a SMART integrated fusion pump that improves drug delivery while reducing clinical workloads and dosing errors. In FIG. 18, graph (a) is a schematic showing the architecture of SMART integrated infusion pumps delivering drugs through a stopcock assembly with a motorized flush valve. All the infusion pumps and the motorized flush valves are interfaced with a computer (e.g., a single board computer) running 27 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) SMART. Image (b) of FIG. 18 is a physical realization of SMART integrated standalone syringe pumps with an intuitive graphical user interface (GUI), and image (c) is a physical realization of SMART integrated standalone peristaltic pumps with an intuitive GUI. FIG.18, graph (d) shows a bench top recreation of clinical practices shows that SMART mediated drug delivery (orange curve) is dramatically superior in delivering drugs accurately and rapidly as compared to existing methods of manually managing infusions (blue curves). FIG. 18, graph (e) shows that SMART can also mitigate fluctuations in drug delivery kinetics associated with drug container (syringe) changes. FIG.18, graph (f) shows that SMART can also significantly improve the drug cessation kinetics in multi-drug infusions, with significantly faster cessation of the intended drug (Drug 1) with minimal fluctuation to the other drug (Drug 2).

[0093] The computational and memory efficient SMART framework can be deployed in a lower footprint single-board computer (SBC) such as the Raspberry Pi. FIG.19 illustrates the performance of a Raspberry Pi implementation. The graph of FIG. 19 shows that the developed algorithms run faithfully on an SBC (designated as PI on the graph) with no statistical performance difference as compared to a high-performance workstation (designated as PC in the graph). Infusion pumps and any associated accessories such as motorized flush valves can be interfaces with SBCs running SMART via, for example, serial communication ports for synchronized control. FIG.20 illustrates an example schematic of such a system.

[0094] FIG. 20 shows a stopcock assembly made up of five commercially available stopcocks, each controlled by a servomotor (in one implementation, Servo Motor MG995360°) capable of independent actuation. As the processor for controlling the servomotors in the illustrated example, an Arduino Uno Rev3 is connected to a power supply and programmed with a code that can read commands sent via serial from the main computer or an SBC (e.g., Raspberry Pi) and translate them into commands for each of the servomotors. An example of how the computer sends a command through the serial port is, in this example, having the four servos (A, B, C, and D) positioned at 180°, 90°, 0°, and 270°, respectively. The flush valve is the last stopcock in the assembly and is connected to an empty reservoir that is used to completely empty the fluid within the stopcock assembly when necessary.

[0095] These low footprint SBCs can be integrated into clinical infusion pumps and interfaced with an intuitive GUI for realizing clinically compatible standalone SMART infusion pumps. Leveraging two infusion technologies used in clinics, two versions of 28 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) standalone SMART infusion pumps were developed: one employing syringe pumps (graph (b) of FIG. 18) and the other employing peristaltic pumps (graph (c) of FIG. 18). The infusion mechanism has a minimal impact on the performance of SMART, as shown in FIG. 21 for single drug (graph (a)) and two drug (graph (b)) control with peristaltic pumps, making this a broadly approachable clinical technology for controlling syringe pumps, peristaltic pumps, or a combination of the two.

[0096] SMART dramatically improves drug infusion kinetics and accuracy as compared to comparative clinical practices, as shown in graph (d) of FIG. 18. Expediting the delivery of drugs in the clinic using the comparative practices relies on one of the two broad manual approaches: continuously making ad-hoc adjustments to the pump flow rates or commencing the drug infusion at a high flow rate before initiating a step change to the set flow rates after an arbitrary period of time (see FIG. 10). As can be seen in graph (d) of FIG. 18, a benchtop recreation of these manual methods reveals the subpar infusion kinetics (t90 as large as 40 minutes) and accuracy (overshoots as large as 200%) in the comparative techniques as compared to SMART (t90 ≈ 7.5 minute and negligible overshoots). Additionally, the comparative techniques require extended interventions by clinicians, significantly increasing their cognitive workload. SMART, on the other hand, requires no manual interventions.

[0097] SMART can also be interfaced with motorized flush valves to enhance procedures accompanying multi-drug infusions such as syringe changes (graph (e) of FIG.18) and drug cessations (graph (f) of FIG. 18). Syringe changes or exchanges are routinely performed by clinicians to replace depleted drug syringes. This procedure transiently interrupts the infusion of the depleted drug, leading to unwanted fluctuations in drug delivery rates that can adversely affect the physiological parameters. Benchtop experiments (see FIG.10) reveal the presence of both instantaneous and delayed fluctuations in drug delivery rates, with the former arising from the instantaneous reduction in flow rate associated with the depleted drug. SMART can entirely or almost entirely eliminate this fluctuation by proportionally altering the flow rate of the other drugs in the system. The interrupted infusion also alters the depleted drug’s concentration within the infusion manifolds, which is perceived as delayed fluctuations when the drug stream is eventually transported to the catheter tip. SMART can reduce or minimize the delayed fluctuations by, for example, rapidly flushing the infusion manifolds with 29 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) the help of motorized flush valves, and resetting all the drug concentrations to those that existed prior to the syringe change.

[0098] Utilizing an analogous procedure, SMART can also rapidly cease the infusion of a selected drug with minimal fluctuations in the delivery of the remaining drugs. In this case, immediately after stopping the infusion of the desired drug, SMART flushes the infusion manifolds so as to reset the concentration to those that should exist for the remaining drugs. This approach reduces t10 (the time to attain 10% of the existing drug delivery rate) by over20 minutes (i.e., ^^^^,^୭ େ୭୬^୰୭୪ െ ^^^^,ୗ^^ୖ^ ^ 20 min), while simultaneously improving thedelivery kinetics of the remaining drugs.

[0099] In anesthesia and critical care settings, multiple life-critical medications formulated in concentrated solutions may be administered by medical infusion pumps via manifolds comprising stopcocks and millimeter-gauge catheters. These multidrug infusions, in comparative implementations, require careful manual management by clinicians including starting, stopping, or dose-adjusting in short time frames while constraining the total fluid delivered to the patient. Experience-based and intuitive manual management of infusion pumps is difficult due to the increasingly complex landscape of multidrug infusions brought about by advancements in surgery, anesthesia and critical care such as modern cardiac, vascular and neurosurgical procedures and also the particularly complex context of neonatal procedures and critical care. The high-stakes infusion management demands are also complicated by the obscured and coupled transport of drugs within infusion manifolds, which can introduce unpredictable delivery delays and mismatches between clinically intended and actual drug delivery profiles. These problems, together with strained health care systems with low clinician-patent ratios, demand alternatives that can enhance medication delivery accuracy while reducing clinical physical and cognitive workloads. Automated systems will have particular utility in remote or resource-constrained environments, such as the military context. The SMART framework set forth herein addresses these and other issues and presents a drug- transport-physics informed approach that improves multidrug infusions through active synchronized control of infusion pumps.

[0100] In the above discussion of SMART, the Shafer number ^^ℎ ൌ 2^^^^^௧^௧ / ^^ௗ^^ௗwas developed, as a non-dimensional number that quantifies the relative magnitude of a drug’s therapeutic action timescale after entering the patient (^^^) to the drug’s transport timescale 30 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)within infusion manifolds (^^^ ൌ ^^ௗ^^ௗ / 2^^௧^௧). When ^^ℎ ^ 1, the drug’s transport timethrottles the effective time for the drug to display its therapeutic effect, potentially leading to medication dose errors and confounding the accurate clinical management of multidrug infusions. SMART overcomes these issues by, among other things, providing automated end- to-end management of multidrug infusions through real-time tracking of drug transport within infusion manifolds, and the use of this information for synchronized control of infusion pumps to attain desired drug delivery endpoints. SMART obtains spatio-temporally resolved drug concentrations within infusion manifolds in real-time by leveraging the Gill- Sankarasubramanian approximation to the advection-diffusion equation with the infusion manifolds approximated through a network model with matched outlet-inlet boundary conditions. This information is leveraged by SMART to provide real-time control of multidrug infusions utilizing an ensemble of deterministic and deep reinforcement learning based decision networks. This reinforcement learning framework for infusion pump mediated drug delivery employs a dual buffer strategy that enhances both training speed and performance. For training the reinforcement learning agent employed by SMART, the “infusion gym” (compatible with the OpenAI gym standard) was constructed, promising easy integration within a variety of deep reinforcement learning workflows.

[0101] The computationally efficient SMART framework deployed in standalone infusion pumps enabled real-time visualization of drug transport within infusion manifolds and significantly improved the drug delivery kinetics while decreasing dosing errors and clinical workloads. In experimental bench top simulations of critical care scenarios, the SMART framework reduced the delivery delays by over 80% and the dosing errors by over 75%, while reducing clinical cognitive workloads by completely eliminating the need for manual control. The ability to visualize and autonomously manage complex multi-drug infusions end-to-end may be used not only for improving anesthesia and critical care management but is also an enabling technology for other clinical situations such as whole organ cryopreservation and cellular recovery after prolonged warm ischemia. SMART has applications for patient care in situations with no or minimal clinical personnel such as in remote, disaster, and conflict zones. Beyond clinical settings, the ability for end-to-end management of fluid delivery has applications in several fields including microfluidic synthesis, LNP-MRNA drug formulation and chemical mixing. 31 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)

[0102] Accordingly, FIG. 20 illustrates one example of a multidrug delivery system 200 in accordance with the present disclosure. The system 200 includes a plurality of infusion pumps 202, an infusion manifold 204, and a controller 206. In FIG. 20, insets are presented showing detail regarding one example of an infusion pump 202 and one example of an infusion manifold 204. The plurality of infusion pumps 202 may include, in some implementations, syringe pumps, peristaltic pumps, or combinations of syringe pumps and peristaltic pumps. However, as noted above, any type of infusion pump or combination of types of infusion pump is within the scope of the present disclosure. The plurality of infusion pumps 202 includes, in examples, a first drug delivery pump configured to output a first drug to the infusion manifold 204 and a carrier fluid pump configured to output a carrier fluid to the infusion manifold 204. The plurality of infusion pumps may include a second drug delivery pump configured to output a second drug to the infusion manifold 204. While FIG.20 illustrates four total infusion pumps 202, in practical implementation any number of infusion pumps 202 (greater than two, or greater than three if one of the infusion pumps 202 provides carrier fluid) may be present in the system 200.

[0103] The infusion manifold 204 includes a plurality of fluid input ports, respective ones are connected to a corresponding infusion pump 202. The infusion manifold 204 further includes a fluid mixing region where, for example, drugs and / or fluids delivered by the infusion pumps 202 mix. An output of the infusion pump 202 may be in fluid communication with a drug delivery device, such as a catheter tip, thereby to deliver a multidrug infusion to a patient or subject. While FIG.20 illustrates an infusion manifold 204 having five fluid input ports, in practical implementations any number of fluid input ports may be present so long as the number is greater than or equal to the number of fluids received by the infusion manifold 204. In the illustrated example, the infusion manifold 204 includes a plurality of stopcocks connected in series. A plurality of actuators may be present, either as part of the infusion manifold 204 or as a separate component, to control the stopcocks. The infusion manifold 204 may include a plurality of motorized flush valves respectively corresponding to the plurality of stopcocks.

[0104] The controller 206 includes at least one processor and a memory. In example implementations, the controller 206 is an SBC (e.g., a Raspberry Pi). The controller 206 is operatively connected to the infusion pumps 202 and the infusion manifold 204. The controller 206 is configured to control an operation of the infusion pumps 202 and / or the infusion 32 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) manifold 204 using the operative connections. The operative connections may be wired connections and / or wireless connections, in any combination. The controller 206 is configured to automatically perform operations, such as any one or more of the operations discussed herein, and thereby implement the SMART framework. In an example, these operations include modeling, in real-time, a spatiotemporal evolution of a drug concentration in the fluid mixing region in accordance with a transport model, and dynamically adjusting a rate of fluid flow from at least one of the plurality of infusion pumps to at least one of the plurality of fluid input ports of the infusion manifold based on the modeling. The controller 206 may further be configured to independently control actuation of the plurality of actuators, thereby to allow for modification of flow between individual infusion pumps 202 and corresponding fluid input ports of the infusion manifold 204. In one example, the controller 206 is configured to independently control the plurality of motorized flush valves to modify (e.g., terminate) flow from one of the plurality of infusion pumps to the fluid mixing region, without altering flow from other ones of the plurality of infusion pumps.

[0105] As described in more detail above, the transport model may include a deterministic drug delivery control policy configured to reduce a difference between an expected drug delivery rate and an actual drug delivery rate at a given time, a trained reinforcement learning model configured to increase a reward at a given time, or combinations thereof. Where a trained reinforcement learning model is utilized, the reward may be based on at least one of an action penalty that penalizes a large action change and an action on a fluid input port that does not receive a drug, an overshoot penalty that penalizes an overshoot of an actual drug delivery rate relative to a target drug delivery rate, a forward reward that rewards an approach of the actual drug delivery rate toward the target drug delivery rate, or a destination reward that rewards a match between the actual drug delivery rate and the target drug delivery rate.

[0106] Infusion pumps integrated with a central processor running a deterministic and machine learning based framework, which can dynamically assemble in memory fluidic paths within drug infusion manifolds and catheters, compute in real time the coupled flow of multiple drugs within the assembled fluidic paths, and control the delivery of drugs by synchronous manipulation of infusion pumps. These pumps can be seamlessly interfaced with each other to facilitate coordinated multidrug infusions. The pumps may also have a user interface, such as 33 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) a graphical user interface (GUI). Two non-limiting examples are expressly contemplated: a pump integrated GUI as well as a standalone GUI (e.g., on a hand-held device or on augmented reality glasses). The controller 206 may be operatively connected to the user interface and may receive at least one of a target delivered excess carrier fluid volume or a target reduction in a dead volume transport delay in support of the operation of dynamically adjusting the rate of fluid flow.

[0107] FIG.22 illustrates an example of a GUI 210 in accordance with various aspects of the present disclosure. In examples, the GUI 210 is an example of the GUI discussed above with regard to images (b) and (c) of FIG.18. The GUI 210 may be presented to an operator via a display device of a computing system having at least one processor and a memory in communication with the at least one processor. The computing system may be, without limitation, a laptop computer, a desktop computer, a notebook computer, a tablet computer, a smartphone, a personal digital assistant (PDA), and the like. In some implementations, the SMART techniques described above may be embodied in the form of a non-transitory computer-readable medium (e.g., in, connected to, or associated with the memory) that stores instructions that, when executed by the at least one processor, cause the computing device to perform various operations relating to presenting the GUI 210 to the user and / or controlling a SMART integrated infusion pump. The computing system may be operatively connected to the multidrug delivery system (e.g., as illustrated in FIG.20) and thus may be referred to herein as a multidrug delivery control device.

[0108] The GUI 210 includes an operating parameter input section 212 which includes a plurality of columns wherein the user may input operating parameters. As illustrated, these include a pump selection column whereby the user may select one of the plurality of infusion pumps, a fluid pump column whereby the user may select a fluid that is administered by the selected infusion pump, and two dose selection columns whereby the user may select a dose for the selected drug in both amount and relative concentration. While FIG. 22 illustrates the operating parameter input section 212 as including drop down selections and text boxes, in practical implementations any form of input (e.g., drop down selections, checkboxes, radio buttons, sliders, dials, etc.) may be utilized. Moreover, while FIG.22 illustrates four rows (thus providing the ability to input parameters for four different pumps simultaneously), in practical implementations any number of rows may be provided. 34 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)

[0109] The GUI 210 also includes a constraint parameter input section 214 to permit the user to determine a tradeoff between the delivered excess carrier fluid volume and the reduction in the dead volume transport delay. Section 214, as illustrated, a first GUI element to permit the user to input a target tradeoff and a second GUI element to display to the user to the corresponding surplus volume. In one example, the slider may be set to 0, implying there is no tradeoff. The maximum reduction in dead volume delay is obtained with the least possible excess carrier fluid that guarantees performance. In another example, the slider is set to 0.5 for a patient that cannot handle excess fluid. A moderate reduction in dead volume delay is obtained, albeit with a smaller excess carrier fluid volume. As above, FIG.22 shows only one example of a possible input form (i.e., a slider). In other implementations, other types of visualizations may be provided.

[0110] The GUI 210 further includes a drug delivery profile visualization section 216. Section 216 may be presented to the user to visualize an expected or actual temporal profile of drug delivery corresponding to the selected input parameters (i.e., via sections 212 and 214). In one example, the visualization presented on section 216 may correspond to a prediction of the spatiotemporal evolution of the drug concentration. In this manner, a user may be able to preview the expected profile and adjust one or more of the input parameters as desired. In another example, the visualization presented on section 216 may correspond to a real-time display of the spatiotemporal evolution of the drug concentration during multidrug infusion. The GUI 210 may be configured to present either or both the preview and the real-time display depending on whether the user has initiated the infusion operation.

[0111] Thus, the GUI 210 may permit the user to interface with the multidrug delivery control device. The multidrug delivery control device may therefore be configured with package of a software, firmware, hardware, or combination thereof that cause the control device to perform various operations. These operations include presenting the GUI 210 to the user via a display of the control device. These operations also include receiving, via the GUI 210, a first user input corresponding to an input operating parameter (e.g., via section 212) and a second user input corresponding to an input constraint (e.g., via section 214). Based on the first and second user inputs, the operations may include displaying, via the GUI 210, a visualization of the drug delivery profile (e.g., via section 216). 35 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694)

[0112] The GUI 210 may also include interface elements by which the user may input a command to begin the multidrug delivery operation. In response to the command, the control device may instruct a multidrug delivery system operating under the control of the control device to perform operations, including but not limiting to real-time modeling of the spatiotemporal evolution of the drug concentration in a fluid mixing region of an infusion manifold in accordance with a transport model, and dynamically adjust a rate of fluid flow from at least one of a plurality of infusion pumps to at least one of a plurality of fluid input ports of the infusion manifold based on the modeling.

[0113] FIG. 23 illustrates one example of a multidrug delivery management method 212 in accordance with the present disclosure. For purposes of illustration and explanation, the method 220 will be described as being performed by, in, or under the control of the controller 206. However, it should be understood that the method 220 may be performed by, in, or under the control of another controller that includes at least one electronic processor and a memory.

[0114] As illustrated, the method 220 includes an operation 222 of receiving a 3D internal fluid volume of an infusion manifold (e.g., the infusion manifold 204) that includes a plurality of fluid input ports in fluid communication with a plurality of infusion pumps (e.g., the infusion pumps 202). The 3D infusion fluid volume may be obtained by any of the above- described techniques (e.g., microCT scanning). At operation 224, the method 220 includes modeling, in real-time, a spatiotemporal evolution of a drug concentration in the fluid mixing region in accordance with the transport model. The drug concentration may correspond to a plurality of drugs in a carrier fluid. Operation 224 may include determining an approximation (e.g., a Gill-Sankarasubramanian approximation) of an advection-diffusion equation for fluids within the mixing region.

[0115] In examples, the transport model may include a deterministic drug delivery control policy configured to reduce a difference between an expected drug delivery rate and an actual drug delivery rate at a given time, or a trained reinforcement learning model configured to increase a reward at a given time. The trained reinforcement learning model may utilize an output of the deterministic drug delivery control policy. In one particular example, the trained reinforcement learning model is based on a dual buffer including a persistent buffer with trajectories obtained from a deterministic model and a first-in-first-out experience buffer. Where a trained reinforcement learning model is used, the reward may be based on at least one 36 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) of an action penalty that penalizes a large action change and an action on a fluid input port that does not receive a drug, an overshoot penalty that penalizes an overshoot of an actual drug delivery rate relative to a target drug delivery rate, a forward reward that rewards an approach of the actual drug delivery rate toward the target drug delivery rate, or a destination reward that rewards a match between the actual drug delivery rate and the target drug delivery rate.

[0116] At operation 226, a rate of fluid flow from at least one of the plurality of infusion pumps to at least one of the plurality of fluid input ports of the infusion manifold is dynamically adjusted based on the modeling of operation 224. The rate of fluid flow may be adjusted by controlling an output of one or more of the infusion pumps and / or by controlling one or more of the plurality of fluid input ports of the infusion manifold. In some examples, the method 220 further includes an operation 228 of delivering a multidrug infusion (e.g., a plurality of drugs) to a patient or subject in accordance with the dynamically adjusted fluid flow rate of operation 226. As noted above, the method 220 may be performed through the use of a user interface. In such examples, the method 220 may include receiving a user input via the user input, the user input indicates at least one of a target delivered excess carrier fluid volume or a target reduction in a dead volume transport delay (e.g., in view of the tradeoff discussed above). Then, operation 226 may be performed based on the user input.

[0117] In one example of operation, for minimizing multi drug dead volume delivery delays, the clinician interfaces a required number of specialized infusion pumps for the desired multidrug infusion. Subsequently, the clinician programs the GUI with the desired drug names, concentrations, and desired infusion rates along with details of the infusion manifolds and other constraints. The infusion is started through the GUI. After reading the specified parameters and the constraints, the device computationally assembles the fluid path within the connected stopcocks and catheters (infusion manifolds) and solves the spatiotemporally resolved flow of the drug. The device then employs a constrained multi-objective drug sequence independent optimization algorithm to compute the instantaneous drug and carrier infusion rates that minimizes the difference between the current and intended drug delivery rates. The device enforces these rates on the connected infusion pumps and simultaneously updates the previously computed spatiotemporally resolved drug distribution with the new infusion rates. Utilizing the updated drug distribution information, the optimization routine is again invoked to compute the best drug and carrier infusion rates that minimizes the difference between the 37 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) updated and intended drug delivery rates. This process is repeated until there is no difference between the actual and intended drug delivery rates. In a bench top simulation of pediatric critical care, where every second is important, the above framework reduced the dead volume delay in drugs entering in the body from over 30 minutes to less than 5 minutes. Overall, the above-described system has the potential to dramatically change multi-drug infusion management in critical care settings, enhancing patient safety and reducing health care provider workloads.

[0118] Other examples and uses of the disclosed technology will be apparent to those having ordinary skill in the art upon consideration of the specification and practice of the invention disclosed herein. The specification and examples given should be considered exemplary only, and it is contemplated that the appended claims will cover any other such embodiments or modifications as fall within the true scope of the invention.

[0119] The Abstract accompanying this specification is provided to enable the United States Patent and Trademark Office and the public generally to determine quickly from a cursory inspection the nature and gist of the technical disclosure and in no way intended for defining, determining, or limiting the present invention or any of its embodiments. 38 QB\93858893.4

Claims

Docket No. MGH 2024-047-03 (125141.04694) CLAIMS What is claimed is:

1. A multidrug delivery system comprising: a plurality of infusion pumps; an infusion manifold including a plurality of fluid input ports respectively connected to the plurality of infusion pumps and including a fluid mixing region; and a controller including at least one processor and a memory, the controller being operatively connected to the plurality of infusion pumps and the infusion manifold and being configured to automatically perform operations comprising: modeling, in real-time, a spatiotemporal evolution of a drug concentration in the fluid mixing region in accordance with a transport model, and dynamically adjusting a rate of fluid flow from at least one of the plurality of infusion pumps to at least one of the plurality of fluid input ports of the infusion manifold based on the modeling.

2. The system of claim 1, wherein the plurality of infusion pumps includes at least one of a syringe pump or a peristaltic pump.

3. The system of claim 1, wherein the plurality of infusion pumps includes a first drug delivery pump configured to output a first drug to the infusion manifold, and a carrier fluid delivery pump configured to output a carrier fluid to the infusion manifold.

4. The system of claim 3, wherein the plurality of infusion pumps further includes a second drug delivery pump configured to output a second drug to the infusion manifold.

5. The system of claim 1, wherein the transport model includes a deterministic drug delivery control policy configured to reduce a difference between an expected drug delivery rate and an actual drug delivery rate at a given time. 39 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) 6. The system of claim 1, wherein the transport model includes a trained reinforcement learning model configured to increase a reward at a given time.

7. The system of claim 6, wherein the reward is based on at least one of an action penalty that penalizes a large action change and an action on one of the plurality of fluid input ports that does not receive a drug, an overshoot penalty that penalizes an overshoot of an actual drug delivery rate relative to a target drug delivery rate, a forward reward that rewards an approach of the actual drug delivery rate toward the target drug delivery rate, or a destination reward that rewards a match between the actual drug delivery rate and the target drug delivery rate.

8. The system of claim 1, wherein the operation of dynamically adjusting the rate of fluid flow is based on at least one of a target delivered excess carrier fluid volume or a target reduction in a dead volume transport delay.

9. The system of claim 1, wherein the controller is operatively connected to a user interface, and wherein the at least one of the target delivered excess carrier fluid volume or the target reduction in the dead volume transport delay is received from a user via the user interface.

10. The system of claim 1, wherein the controller is a single-board computer.

11. The system of claim 1, wherein the infusion manifold includes a plurality of stopcocks connected in series.

12. The system of claim 11, further comprising a plurality of actuators respectively configured to control the plurality of stopcocks, wherein the controller is configured to automatically perform operations further comprising independently controlling actuation of the plurality of actuators.

13. The system of claim 11, wherein the infusion manifold includes a plurality of motorized flush valves respectively corresponding to the plurality of stopcocks, wherein each 40 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) of the plurality of motorized flush valves are configured to be independently controlled by the controller to terminate flow from one of the plurality of infusion pumps without altering flow from others of the plurality of infusion pumps.

14. A multidrug delivery management method comprising: receiving a three-dimensional internal fluid volume of an infusion manifold, the infusion manifold including a plurality of fluid input ports respectively connected to a plurality of infusion pumps and including a fluid mixing region; modeling, in real-time, a spatiotemporal evolution of a drug concentration in the fluid mixing region in accordance with the transport model; and dynamically adjusting a rate of fluid flow from at least one of the plurality of infusion pumps to at least one of the plurality of fluid input ports of the infusion manifold based on the modeling.

15. The method of claim 14, wherein the transport model includes a deterministic drug delivery control policy configured to reduce a difference between an expected drug delivery rate and an actual drug delivery rate at a given time.

16. The method of claim 14, wherein the transport model includes a trained reinforcement learning model configured to increase a reward at a given time.

17. The method of claim 16, wherein the reward is based on at least one of an action penalty that penalizes a large action change and an action on one of the plurality of fluid input ports that does not receive a drug, an overshoot penalty that penalizes an overshoot of an actual drug delivery rate relative to a target drug delivery rate, a forward reward that rewards an approach of the actual drug delivery rate toward the target drug delivery rate, or a destination reward that rewards a match between the actual drug delivery rate and the target drug delivery rate.

18. The method of claim 16, wherein the trained reinforcement learning model is based on a dual buffer including a persistent buffer with trajectories obtained from a deterministic model and a first-in-first-out experience buffer. 41 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) 19. The method of claim 14, wherein the operation of modeling includes determining an approximation of an advection-diffusion of fluids within the fluid mixing region.

20. The method of claim 19, wherein the approximation is a Gill- Sankarasubramanian approximation.

21. The method of claim 14, further comprising: delivering a plurality of drugs via the infusion manifold in accordance with the dynamically adjusted rate of fluid flow.

22. The method of claim 14, further comprising: receiving a user input via a user interface, wherein the user input indicates at least one of a target delivered excess carrier fluid volume or a target reduction in a dead volume transport delay.

23. The method of claim 22, wherein the operation of dynamically adjusting the rate of fluid flow is based on the user input.

24. A non-transitory computer-readable medium storing instructions that, when executed by at least one processor of a multidrug delivery control device, cause the control device to perform operations comprising: presenting a graphical user interface (GUI) to a user via a display of the control device; receiving, via the GUI, a first user input corresponding to an input operating parameter and a second user input corresponding to an input constraint; displaying, via the GUI, a visualization of a drug delivery profile based on the first user input and the second user input; and in response to a command to begin a multidrug delivery operation, instruct a multidrug delivery system to perform operations including: modeling, in real-time, a spatiotemporal evolution of a drug concentration in a fluid mixing region of an infusion manifold in accordance with a transport model, and 42 QB\93858893.4Docket No. MGH 2024-047-03 (125141.04694) dynamically adjust a rate of fluid flow from at least one of a plurality of infusion pumps to at least one of a plurality of fluid input ports of the infusion manifold based on the modeling.

25. The non-transitory computer-readable medium of claim 24, wherein the visualization of the drug delivery profile corresponds to a prediction of the spatiotemporal evolution of the drug concentration.

26. The non-transitory computer-readable medium of claim 24, wherein the visualization of the drug delivery profile corresponds to a real-time display of the spatiotemporal evolution of the drug concentration. 43 QB\93858893.4

Images