Abstract: A machine learning model, including: a categorical input feature, having a defined set of values; a plurality of non-categorical input features; a word embedding layer configured to convert the categorical input feature into an output in a word space having two dimensions; and a machine learning network configured to receive the output of the word embedding layer and the plurality of non-categorical input features and to produce a machine learning model output.

Abstract: Systems, apparatuses, and methods provide for the monitoring and/or management of acute kidney injury (AKI). For example, an apparatus (100) is configured to determine whether a baseline AKI risk prediction is above a baseline threshold based on patient demographic data and patient medical history data, and perform a continuous AKI risk prediction. The continuous AKI risk prediction includes determining whether an any risk of AKI prediction is above an any AKI threshold based on dynamic intervention data and/or dynamic patient condition data, and determining an AKI stage prediction in response to a determination that the any risk of AKI prediction is above the any AKI threshold based on the dynamic intervention data and/or the dynamic patient condition data.

Abstract: In a risk level assessment method for a plurality of clinical conditions as follows, a set of laboratory test results (32) are stored with time stamps for a patient, including at least one hematology test result and at least one arterial blood gas (ABG) test result. For each clinical condition, a risk level is determined for the clinical condition based on a clinical condition-specific sub-set of the stored set of laboratory test results. This determination is made conditional on the stored clinical condition-specific sub-set of laboratory test results being sufficient to determine the risk level. A time stamp is assigned to each determined risk level based on the time stamps for the laboratory test results of the clinical condition-specific sub-set of laboratory test results. A display device (44, 46) displays the determined risk level and the assigned time stamp for each clinical condition whose determined risk level satisfies a display criterion.

Abstract: A method and system for generating real time workload balancing recommendations comprising receiving transition data, medical data, and staffing data; determining a transition probability for each of a plurality of patients; determining a predicted workload to be generated by each of the plurality of patients; simulating the predicted workload to be generated by each of the plurality of patients, the future workload for each of a plurality of units in the hospital; generating staffing recommendations; and displaying the generated staffing recommendations on a user display of the workload balancing system.

Abstract: A method for performing, using a patient disposition system, a disposition analysis of a plurality of patients to optimize a discharge planning process for each of the plurality of patients, including: (i) receiving electronic medical record information about each of the plurality of patients; (ii) identifying one of a plurality of different patient types for each of the plurality of patients based on the received electronic medical record information; (iii) selecting a trained multi-state model for each identified patient type; and (iv) determining, based on the selected trained multi-state model, a disposition state for each of the plurality of patients in real-time, wherein each disposition state includes a location to which the patient is to be discharged. The method further includes determining at least one service or assessment that can be deferred to the location to which the patient is to be discharged.

Abstract: Methods and systems for determining the blood pressure of a patient. The system may include a first device configured to collect a first plurality of blood pressure measurements of the patient, a second device configured to collect a second plurality of blood pressure measurements of the patient, and a processor configured to identify a divergence between the first plurality and the second plurality, retrieve, from a memory, a clinical event, compare the first plurality and the second plurality to the clinical event, and determine that the first plurality is more accurate than the second plurality based on the comparison.

Abstract: A method for training a baseline risk model, including: pre-processing input data by normalizing continuous variable inputs and producing one-hot input features for categorical variables; providing definitions for clean input data and dirty input data based upon various input data related to a patient condition; segmenting the input data into clean input data and dirty input data, wherein the clean input data includes a first subset and a second subset, where the first subset and the second subset include all of the clean input data and are disjoint; training a machine learning model using the first subset of the clean data; and evaluating the performance of the trained machine learning model using the second subset of the clean input data and the dirty input data.

Abstract: Techniques are described herein for using time series data such as vital signs data and laboratory data or other time series data as input across machine learning models to predict a change in stage of a medical condition of a patient. In various embodiments, patient data comprising vital signs data of a patient and laboratory data or other time series data of the patient corresponding to an observation window may be received. A time series model may be used to predict a change in stage of a medical condition in the patient in a prediction window based on the patient data. The predicted change in stage of the medical condition may be output.

Abstract: Various embodiments described herein relate to methods and apparatus for robust classification. Many real-world datasets suffer from missing or incomplete data. By assigning weights to certain features of a dataset based on which feature(s) are missing or incomplete, embodiments of the prevention can provide robustness and resilience to missing data.

Abstract: Systems and methods for adapting a first machine learning model that takes clinical data as input, based on a second set of training data. The first machine learning model having been trained on a first set of training data. The method comprises adding an adaption module to the first machine learning model, the adaption module comprising a second machine learning model, and training the second machine learning model using a second set of training data to take an output of the first machine learning model as input and provide an adjusted output.

Abstract: A system (100) for assessing fluid responsiveness includes an infusion pump (24) in communication with at least one processor (32), and a plurality of physiological monitors (40,42,44,46) operable to receive physiological signals from an associated patient. Physiological signals (48,50) acquired from the associated patient (10) during a fluid challenge are synchronized with a timing signal (54) of the infusion pump (24) administering the fluid challenge. One or more dynamic indices and/or features (58) is calculated from the synchronized physiological signals (50), and one or more dynamic indices and/or features (50) is calculated from baseline physiological signals (48) acquired from the associated patient (10) prior to the fluid challenge. A fluid responsiveness probability value (64) of the patient (10) is determined based on dynamic indices and/or features (58) from the synchronized physiological signals (50) and dynamic indices and/or features (50) from the baseline physiological signals (48).

Abstract: A machine learning model, including: a categorical input feature, having a defined set of values; a plurality of non-categorical input features; a word embedding layer configured to convert the categorical input feature into an output in a word space having two dimensions; and a machine learning network configured to receive the output of the word embedding layer and the plurality of non-categorical input features and to produce a machine learning model output.

Abstract: A method and system method for determining a patient's acute kidney injury (AKI) stage, including: receiving a patient's current AKI stage; calculating a patient's new AKI stage; comparing the new AKI stage to the current AKI stage; updating the patient's AKI stage to the new AKI stage when the new AKI stage is greater than the current AKI stage; calculating AKI stage exit criteria and an AKI exit stage value; determining whether the AKI stage exit criteria are satisfied; and reducing the patient's AKI stage to the exit AKI stage when the AKI stage exit criteria are satisfied.

Abstract: A method for training a baseline risk model, including: pre-processing input data by normalizing continuous variable inputs and producing one-hot input features for categorical variables; providing definitions for clean input data and dirty input data based upon various input data related to a patient condition; segmenting the input data into clean input data and dirty input data, wherein the clean input data includes a first subset and a second subset, where the first subset and the second subset include all of the clean input data and are disjoint; training a machine learning model using the first subset of the clean data; and evaluating the performance of the trained machine learning model using the second subset of the clean input data and the dirty input data.

Abstract: Methods and systems for monitoring patient physiological status. The system may include a source of vital sign measurements for a patient, a trained machine learning model that receives the vital sign measurements and provides an output related to the physiological status of the patient, and an interface configured to present the output to an operator. The method may include receiving, at a trained machine learning model, at least one physiological measurement, demographic information point, or treatment plan for a patient, providing, using the trained machine learning model, an output relating to the physiological status of the patient, and presenting, using an interface, the output to an operator.

Abstract: The following relates generally to the medical monitoring arts, medical warning systems concerning a monitored patient, and so forth. In clinical settings, alarms are usually triggered when a single-parameter or a multi-parameter score exceeds certain thresholds. When a score needs to be determined, if certain parameters are not available, the common practice is to use the most recent measurements of the parameters for the score calculation. However, a patient's status may change from moment to moment. The parameters measured hours ago may not be a good indicator of the patient's current status. This uncertainty can put deteriorating patients at great risk. An embodiment uses statistical methods to estimate a range of scores and the probability of these scores if old measurements have to be used for score determination. Instead of giving a single number at a time, a confidence interval may be displayed to emphasize the fact that the score is determined partially based on old measurements.

Abstract: The present disclosure relates to estimation and use of clinician assessment of patient acuity. In various embodiments, a plurality of patient feature vectors associated with a plurality of respective patients may be obtained (302, 304). Each patient feature vector may include one or more health indicator features indicative of observable health indicators of a patient, and one or more treatment features indicative of characteristics of treatment provided to the patient. A machine learning model (216) may be trained (306) based on the patient feature vectors to receive, as input, subsequent patient feature vectors, and to provide, as output, indications of levels of clinician acuity assessment. Later, a patient feature vector associated with a given patient may be provided (404) as input to the machine learning model. Based on output from the machine learning model, a level of clinician acuity assessment associated with the given patient may be estimated (406) and used (408-416) for various applications.

Abstract: A medical modeling system and method predict a risk of a physiological condition, such as mortality, for a patient. Measurements of a plurality of predictive variables for the patient are received. The plurality of predictive variables are predictive of the risk of the physiological condition. The risk of the physiological condition is calculated by applying the received measurements to at least one model modeling the risk of the physiological condition using the plurality of predictive variables. The at least one model includes at least one of a hidden Markov model and a logistic regression model. An indication of the risk of the physiological condition is output to a clinician.

Abstract: In a risk level assessment method for a plurality of clinical conditions as follows, a set of laboratory test results (32) are stored with time stamps for a patient, including at least one hematology test result and at least one arterial blood gas (ABG) test result. For each clinical condition, a risk level is determined for the clinical condition based on a clinical condition-specific sub-set of the stored set of laboratory test results. This determination is made conditional on the stored clinical condition-specific sub-set of laboratory test results being sufficient to determine the risk level. A time stamp is assigned to each determined risk level based on the time stamps for the laboratory test results of the clinical condition-specific sub-set of laboratory test results. A display device (44, 46) displays the determined risk level and the assigned time stamp for each clinical condition whose determined risk level satisfies a display criterion.

Abstract: Various embodiments described herein relate to methods and apparatus for robust classification. Many real-world datasets suffer from missing or incomplete data. By assigning weights to certain features of a dataset based on which feature(s) are missing or incomplete, embodiments of the prevention can provide robustness and resilience to missing data.