How to assess the risks associated with the usage of a medical device based on predictive modeling: the case of an anemia control model certified as medical device.

BACKGROUND: The successful application of Machine Learning (ML) to many clinical problems can lead to its implementation as a medical device (MD), which is important to assess the associated risks. METHODS: An anemia control model (ACM), certified as MD, may face adverse events as a result of wrong predictions that are translated into suggestions of doses of erythropoietic stimulating agents to dialysis patients. Risks are assessed as the combination of severity and probability of a given hazard. While severities are typically assessed by clinicians, probabilities are tightly related to the performance of the predictive model. RESULTS: A postmarketing data set formed by all adult patients registered in French, Portuguese, and Spanish clinics, belonging to an international network, was considered; 3876 patients and 11,508 suggestions were eventually included. The achieved results show that there are no statistical differences between the probabilities of adverse events that are estimated in the ACM test set (using only Spanish clinics) and those actually observed in the postmarketing cohort. CONCLUSIONS: The risks of an ACM-MD can be accurately and robustly estimated, thus enhancing patients' safety. The proposed methodology is applicable to other clinical decisions based on predictive models since our proposal does not depend on the particular predictive model.

Enhanced prediction of hemoglobin concentration in a very large cohort of hemodialysis patients by means of deep recurrent neural networks.

Erythropoiesis Stimulating Agents (ESAs) have become a standard anemia management tool for End Stage Renal Disease (ESRD) patients. However, dose optimization constitutes an extremely challenging task due to huge inter and intra-patient variability in the responses to ESA administration. Current data-based approaches to anemia control focus on learning accurate hemoglobin prediction models, which can be later utilized for testing competing treatment choices and choosing the optimal one. These methods, despite being proven effective in practice, present several shortcomings which this paper intends to tackle. Namely, they are limited to a small cohort of patients and, even then, they fail to provide suggestions when some strict requirements are not met (such as having a three month history prior to the prediction). Here, recurrent neural networks (RNNs) are used to model whole patient histories, providing predictions at every time step since the very first day. Furthermore, an unprecedented amount of data (∼110,000 patients from many different medical centers in twelve countries, without exclusion criteria) was used to train it, thus allowing it to generalize for every single patient. The resulting model outperforms state-of-the-art Hemoglobin prediction, providing excellent results even when tested on a prospective dataset. Simultaneously, it allows to bring the benefits of algorithmic anemia control to a very large group of patients.

Development of an Artificial Intelligence Model to Guide the Management of Blood Pressure, Fluid Volume, and Dialysis Dose in End-Stage Kidney Disease Patients: Proof of Concept and First Clinical Assessment.

BACKGROUND: Fluid volume and blood pressure (BP) management are crucial endpoints for end-stage kidney disease patients. BP control in clinical practice mainly relies on reducing extracellular fluid volume overload by diminishing targeted postdialysis weight. This approach exposes dialysis patients to intradialytic hypotensive episodes. SUMMARY: Both chronic hypertension and intradialytic hypotension lead to adverse long-term outcomes. Achieving the optimal trade-off between adequate fluid removal and the risk of intradialytic adverse events is a complex task in clinical practice given the multiple patient-related and dialysis-related factors affecting the hemodynamic response to treatment. State-of-the-art artificial intelligence has been adopted in other complex decision-making tasks for dialysis patients and may help personalize the multiple dialysis-related prescriptions affecting patients' intradialytic hemodynamics. As a proof of concept, we developed a multiple-endpoint model predicting session-specific Kt/V, fluid volume removal, heart rate, and BP based on patient characteristics, historic hemodynamic responses, and dialysis-related prescriptions. KEY MESSAGES: The accuracy and precision of this preliminary model is extremely encouraging. Such analytic tools may be used to anticipate patients' reactions through simulation so that the best strategy can be chosen based on clinical judgment or formal utility functions.

An international observational study suggests that artificial intelligence for clinical decision support optimizes anemia management in hemodialysis patients.

Managing anemia in hemodialysis patients can be challenging because of competing therapeutic targets and individual variability. Because therapy recommendations provided by a decision support system can benefit both patients and doctors, we evaluated the impact of an artificial intelligence decision support system, the Anemia Control Model (ACM), on anemia outcomes. Based on patient profiles, the ACM was built to recommend suitable erythropoietic-stimulating agent doses. Our retrospective study consisted of a 12-month control phase (standard anemia care), followed by a 12-month observation phase (ACM-guided care) encompassing 752 patients undergoing hemodialysis therapy in 3 NephroCare clinics located in separate countries. The percentage of hemoglobin values on target, the median darbepoetin dose, and individual hemoglobin fluctuation (estimated from the intrapatient hemoglobin standard deviation) were deemed primary outcomes. In the observation phase, median darbepoetin consumption significantly decreased from 0.63 to 0.46 µg/kg/month, whereas on-target hemoglobin values significantly increased from 70.6% to 76.6%, reaching 83.2% when the ACM suggestions were implemented. Moreover, ACM introduction led to a significant decrease in hemoglobin fluctuation (intrapatient standard deviation decreased from 0.95 g/dl to 0.83 g/dl). Thus, ACM support helped improve anemia outcomes of hemodialysis patients, minimizing erythropoietic-stimulating agent use with the potential to reduce the cost of treatment.

Performance of a Predictive Model for Long-Term Hemoglobin Response to Darbepoetin and Iron Administration in a Large Cohort of Hemodialysis Patients.

Anemia management, based on erythropoiesis stimulating agents (ESA) and iron supplementation, has become an increasingly challenging problem in hemodialysis patients. Maintaining hemodialysis patients within narrow hemoglobin targets, preventing cycling outside target, and reducing ESA dosing to prevent adverse outcomes requires considerable attention from caregivers. Anticipation of the long-term response (i.e. at 3 months) to the ESA/iron therapy would be of fundamental importance for planning a successful treatment strategy. To this end, we developed a predictive model designed to support decision-making regarding anemia management in hemodialysis (HD) patients treated in center. An Artificial Neural Network (ANN) algorithm for predicting hemoglobin concentrations three months into the future was developed and evaluated in a retrospective study on a sample population of 1558 HD patients treated with intravenous (IV) darbepoetin alfa, and IV iron (sucrose or gluconate). Model inputs were the last 90 days of patients' medical history and the subsequent 90 days of darbepoetin/iron prescription. Our model was able to predict individual variation of hemoglobin concentration 3 months in the future with a Mean Absolute Error (MAE) of 0.75 g/dL. Error analysis showed a narrow Gaussian distribution centered in 0 g/dL; a root cause analysis identified intercurrent and/or unpredictable events associated with hospitalization, blood transfusion, and laboratory error or misreported hemoglobin values as the main reasons for large discrepancy between predicted versus observed hemoglobin values. Our ANN predictive model offers a simple and reliable tool applicable in daily clinical practice for predicting the long-term response to ESA/iron therapy of HD patients.

Clustering the lexicon in the brain: a meta-analysis of the neurofunctional evidence on noun and verb processing.

Although it is widely accepted that nouns and verbs are functionally independent linguistic entities, it is less clear whether their processing recruits different brain areas. This issue is particularly relevant for those theories of lexical semantics (and, more in general, of cognition) that suggest the embodiment of abstract concepts, i.e., based strongly on perceptual and motoric representations. This paper presents a formal meta-analysis of the neuroimaging evidence on noun and verb processing in order to address this dichotomy more effectively at the anatomical level. We used a hierarchical clustering algorithm that grouped fMRI/PET activation peaks solely on the basis of spatial proximity. Cluster specificity for grammatical class was then tested on the basis of the noun-verb distribution of the activation peaks included in each cluster. Thirty-two clusters were identified: three were associated with nouns across different tasks (in the right inferior temporal gyrus, the left angular gyrus, and the left inferior parietal gyrus); one with verbs across different tasks (in the posterior part of the right middle temporal gyrus); and three showed verb specificity in some tasks and noun specificity in others (in the left and right inferior frontal gyrus and the left insula). These results do not support the popular tenets that verb processing is predominantly based in the left frontal cortex and noun processing relies specifically on temporal regions; nor do they support the idea that verb lexical-semantic representations are heavily based on embodied motoric information. Our findings suggest instead that the cerebral circuits deputed to noun and verb processing lie in close spatial proximity in a wide network including frontal, parietal, and temporal regions. The data also indicate a predominant-but not exclusive-left lateralization of the network.

Computational intelligence for the Balanced Scorecard: studying performance trends of hemodialysis clinics.

OBJECTIVES: The Balanced Scorecard (BSC) is a general, widely employed instrument for enterprise performance monitoring based on the periodic assessment of strategic Key Performance Indicators that are scored against preset targets. The BSC is currently employed as an effective management support tool within Fresenius Medical Care (FME) and is routinely analyzed via standard statistical methods. More recently, the application of computational intelligence techniques (namely, self-organizing maps) to BSC data has been proposed as a way to enhance the quantity and quality of information that can be extracted from it. In this work, additional methods are presented to analyze the evolution of clinic performance over time. METHODS: Performance evolution is studied at the single-clinic level by computing two complementary indexes that measure the proportion of time spent within performance clusters and improving/worsening trends. Self-organizing maps are used in conjunction with these indexes to identify the specific drivers of the observed performance. The performance evolution for groups of clinics is modeled under a probabilistic framework by resorting to Markov chain properties. These allow a study of the probability of transitioning between performance clusters as time progresses for the identification of the performance level that is expected to become dominant over time. RESULTS: We show the potential of the proposed methods through illustrative results derived from the analysis of BSC data of 109 FME clinics in three countries. We were able to identify the performance drivers for specific groups of clinics and to distinguish between countries whose performances are likely to improve from those where a decline in performance might be expected. According to the stationary distribution of the Markov chain, the expected trend is best in Turkey (where the highest performance cluster has the highest probability, P=0.46), followed by Portugal (where the second best performance cluster dominates, with P=0.50), and finally Italy (where the second best performance cluster has P=0.34). CONCLUSION: These results highlight the ability of the proposed methods to extract insights about performance trends that cannot be easily extrapolated using standard analyses and that are valuable in directing management strategies within a continuous quality improvement policy.

A novel approach to the problem of non-uniqueness of the solution in hierarchical clustering.

The existence of multiple solutions in clustering, and in hierarchical clustering in particular, is often ignored in practical applications. However, this is a non-trivial problem, as different data orderings can result in different cluster sets that, in turns, may lead to different interpretations of the same data. The method presented here offers a solution to this issue. It is based on the definition of an equivalence relation over dendrograms that allows developing all and only the significantly different dendrograms for the same dataset, thus reducing the computational complexity to polynomial from the exponential obtained when all possible dendrograms are considered. Experimental results in the neuroimaging and bioinformatics domains show the effectiveness of the proposed method.

Use of Self-Organizing Maps for Balanced Scorecard analysis to monitor the performance of dialysis clinic chains.

The Balanced Scorecard (BSC) is a validated tool to monitor enterprise performances against specific objectives. Through the choice and the evaluation of strategic Key Performance Indicators (KPIs), it provides a measure of the past company's outcome and allows planning future managerial strategies. The Fresenius Medical Care (FME) BSC makes use of 30 KPIs for a continuous quality improvement strategy within its dialysis clinics. Each KPI is monthly associated to a score that summarizes the clinic efficiency for that month. Standard statistical methods are currently used to analyze the BSC data and to give a comprehensive view of the corporate improvements to the top management. We herein propose the Self-Organizing Maps (SOMs) as an innovative approach to extrapolate information from the FME BSC data and to present it in an easy-readable informative form. A SOM is a computational technique that allows projecting high-dimensional datasets to a two-dimensional space (map), thus providing a compressed representation. The SOM unsupervised (self-organizing) training procedure results in a map that preserves similarity relations existing in the original dataset; in this way, the information contained in the high-dimensional space can be more easily visualized and understood. The present work demonstrates the effectiveness of the SOM approach in extracting useful information from the 30-dimensional BSC dataset: indeed, SOMs enabled both to highlight expected relationships between the KPIs and to uncover results not predictable with traditional analyses. Hence we suggest SOMs as a reliable complementary approach to the standard methods for BSC interpretation.

Interacting with an artificial partner: modeling the role of emotional aspects.

In this paper we introduce a simple model based on probabilistic finite state automata to describe an emotional interaction between a robot and a human user, or between simulated agents. Based on the agent's personality, attitude, and nature, and on the emotional inputs it receives, the model will determine the next emotional state displayed by the agent itself. The probabilistic and time-varying nature of the model yields rich and dynamic interactions, and an autonomous adaptation to the interlocutor. In addition, a reinforcement learning technique is applied to have one agent drive its partner's behavior toward desired states. The model may also be used as a tool for behavior analysis, by extracting high probability patterns of interaction and by resorting to the ergodic properties of Markov chains.