Microsoft word - no35 - full paper.doc

Remote monitoring of chronical pathologies
using personalized Markov models
Laurent JEANPIERRE (University Henri Poincaré, Nancy 1) François CHARPILLET (Directeur de recherche INRIA) Campus Scientifique, BP 239 - F54506, Vandoeuvre-lès-Nancy Phone: +33(0)383592095; Fax: +33(0)383278319
Abstract. In this paper, an efficient method
To cope with so much uncertainty, the choice of probabilistic models seems obvious [1]: presented, which is particularly adapted to their main advantages are their noise tolerance long term monitoring of patients. It results and their ability to model ill-known rules of from the collaboration between the LORIA evolution. More precisely, the class of partially (Lorrain Research Laboratory of Computer observable Markov models [2] is well adapted Science and its Applications), and the ALTIR to the diagnosis aspect of the problem, because of their observation process. With these models, well-known algorithms are able to approach is that the physician can customize uncover the hidden state of the model from the each patient’s model. This model is expressed observations. This is an elegant solution to the in medical terms, simplifying the interaction of diagnosis problem, since it only needs a statistical description of the pathologies we are The approach will be illustrated with the remote monitoring of patients suffering from Even more important, each patient is unique: kidney disease. In a two years prospective even if they almost follow the same generic model [3], there are some important variations monitored by our system, while 15 others were hindering the diagnosis process. For example, monitored the classical way. The two groups’ orthostatic hypotension is a common effect of statistics has shown that the system was really dehydration; however, this is not true with beneficent to patients’ health. This experiment cardiac insufficient patients. For aging people, arteries tend to lose their elasticity; thus, enterprise, to promote and develop this system. variations of blood pressure are much different Keywords: Probabilistic model based reasoning, human- computer interaction, telemedicine, remote monitoring Finally, the real trouble is that we cannot determine the real condition of a given patient, 1 General considerations
with very few exceptions. Even the physician The automated monitoring of a patient is a may not be sure of his own diagnosis. For difficult task, because our knowledge of the example, the only certain symptom of hyper dynamics of patient evolution is limited. The hydration is oedemas presence. But it is remote monitoring is even more difficult, already too late: the patient has to go to the because the monitored patient stays in an hospital. This consideration implies that the uncontrolled environment. Thus, there exists model is very hard to define formally. Even an important noise that spoils data, and alters measuring the accuracy of the diagnosis is the patient evolution. For example, a simple rather difficult, since we cannot know what a stress strongly influences blood pressure. All these considerations show clearly that the order to detect hydration problems, as well as interaction with physicians has to be a central ideal weight variations. The model we present point in the system. Currently, the physicians’ in this paper is the heart of this system. own diagnosis is our only way to define what a meaningful diagnosis is. To have a really 2.2 The Diatelic project
working collaboration with them, we chose to Each patient sends his medical data through ensure a strict medical semantic in each part of the Internet daily. These data mainly include the patient weight, his blood pressure, and parameters of his dialysis. Blood pressure is 2 Medical context
measured when the patient is lit, and when he stands up. The difference between these two 2.1 General dialysis considerations
The Diatelic project is a real case study of the monitoring of patients suffering from chronic Dialysis parameters include information about renal insufficiency. In general, this illness is each one of the four dialysis bags a patient due to kidney diseases which hinder its normal uses daily. In particular, the system can use bag concentration, its weight, and the time it was injected. After the dialysis, each bag is More precisely, our patients are treated weighted, and this measure is added to the through continuous ambulatory peritoneal dialysis. This method allows patients to stay at home while being dialysed daily; they just Once a day, each patient uses an Internet have to come and see their nephrologist once a connection to transfer those data to a dedicated month. Between visits, the patient is left alone, server. There, these data are stored in a database, and are compared to the patient profile, along with data from previous days. In this situation, troubles are mainly related to When some anomaly is detected, an alert is hydration problems: dehydration generally sent to the patient, and to the medical team, results in comas, while hyperhydration results which will figure out its own diagnosis. in increased blood pressure, and may even produce oedemas. High blood pressure can Obviously, all of the gathered medical signals damage the whole organism, and generally are available to the physicians, along with the computed diagnosis during the last months. With these data, they can see the patient’s The objective of the treatment is then to evolution, and determine if the system was regulate the patient weight, which is closely right or if the alert was meaningless. When the related to his hydration level. To achieve this patient really is endangered, the treatment may goal, the physicians compute an ideal weight be adapted to overcome this evolution, or the value for each patient. Then, the dialysis patient might be hospitalized if the standard strength is modified to reduce the difference treatment is insufficient. The objective is to between the patient’s weight and his ideal prevent the worsening of patient health, or at weight. The trouble is this ideal weight can least to reduce his recovery duration by also evolve, as the patient grows bigger, or if detecting troubles before they fully develop. he loses weight. In order to enhance these patients’ health 3 The model
condition, each patient should be monitored on 3.1 Markov models
a daily basis, rather than once a month. Given that each nephrologist often has to monitor All of the Markov models are based on the more than one hundred patients, they cannot Markov property: the behaviour of a system monitor all of them each day. Therefore, we only depends on its current state. These proposed a monitoring architecture [4], in models may be declined as three interesting forms for diagnosing: Markov chains, hidden Markov decision processes. Since each model monitoring, there is a particular state that plays is built on the previous one, I will introduce a crucial role: the “healthy” state. This state is them from the simpler to the more complex the ideal patient condition, where nothing is one. Later, I will explain how each of these wrong. From this point, we can derive other states representing deviations from this healthy Markov models are based on a finite set of condition. Thus, those states will include some states, a finite set of actions, and a function of pathology, or some conjunction of pathologies. transition. The former describes the evolution This model of the patient is relatively rough, of the model state, knowing the current action. since it only knows caricatural situations of Partially observable Markov models [2] add a static conditions. However, it is generally new dimension to these models by inserting sufficient for diagnosing simple pathologies. the observation process. This new stochastic In more complex situations, we may split some process hides the model state: it simply emits a state to represent different grades of the health given observation based on the model state. level of the patient. For example, we could The challenge is then to infer the hidden state split the healthy state in “healthy” and “healthy but stressful”, if this notion was important for the diagnosis. This could even be The last model evolution is the partially necessary to ensure that the model satisfies the observable Markov decision processes. They Markov property. For example, the patient contain another function which is used to evolution may be biased if he is stressful. reward the system. Hence, an objective is defined, and the system will have to choose The obvious advantage of this states definition is that each state has a very strong medical meaning. Thus, physicians can interact with Currently, the latest model is not used yet, the model easily. This helps their cooperation since we only consider the diagnosis process. when first designing the model from scratch, However, we could add some rewards if the and when adapting it to a given patient. Next, system was to choose actions to correct the it helps them when they want to interpret the patient state, or simply to get a more accurate diagnosis by asking for complementary tests. 3.2 Medical semantics
associated transition function models all of the influences the patient can receive. Each action important part of the model. In fact, since the influences the state of the model through a objective is to diagnose some pathology, the transition matrix. This is the simplest way to states of the model must be related to the implement a probabilistic function indicating pathologies we want the system to discover. how the patient state should evolve from one Classical approaches model the patient state state to another. From the medical point of from gathered data, and then try and translate view, this transition function is simply the expected evolution of the patient, in response to some treatment (the action). For example, We chose to introduce the medical semantics aspirin should lower the overall temperature of directly in the model to enhance collaboration the patient. However, in some cases, aspirin is with the medical team. Thus, we have defined not sufficient; this implies that the state of the our states from a medical point of view. Next, model is not precise enough, or simply that, for some unknown reason, the result is not the through the gathered data [1]. This approach enhances the physician's comprehension of the system behaviour, and it also simplifies the system itself since no interpretation is needed. transitions is important: it allows for the use of indicating how probable a given value of this dynamics. Moreover, physicians are used to sensor is, knowing that the patient is in a work with statistics and those transitions help precise state. For example, blood pressure coping with the noise inherent to the normal should be lower than the normal value when the patient is dehydrated; however, in few The observation function is the second most cases, it may be normal. This will show as the important characteristic of the model: it following graph, which is centred on the indicates the influence the state of the patient has on the values we can observe. It is the core of the observation process. For example, when a patient suffers from some infection, his ganglia swell. Again, this behaviour is the standard one; there can be exceptions where a patient suffers from infection, and no ganglion swells. It may be related to a factor that has not been modelled into the states; however, it may simply be related to an unknown external Blood pressure when the patient is dehydrated For these reasons, the observation function is Thanks to the strong semantics of states, this also probabilistic. It indicates the probability kind of graph is really easy to interpret. of each possible observation while the patient Moreover, physicians can interact directly with is in a given state. Different states may have this graph to set the correct influence. similar probability for some observation. This Finally, all of these sensors are aggregated into is why the model state is said hidden: in a single observation probability [8] indicating general, with a given set of observations, no how representative the physiological signals one can determine precisely which the real state of the patient is. In our model, this function is based on fuzzy 4 The diagnosis process
sets [6] that allow the model to use continuous The diagnostic is the art of finding out the observations with relatively few parameters. reasons that explain observed symptoms. With Thanks to this fuzzy notation, each parameter our model, and more particularly with the state has an obvious meaning for the medical team. semantics, diagnosing a medical condition is Additionally, this fuzziness allows for a simple equivalent to finding which the hidden state of but expressive way of describing a given state To uncover the hidden part of the model from Actually, each state is defined separately from the observations, the traditional algorithm of the others; additionally, we imposed that data Viterbi [2] is well adapted. Actually, dynamic from different sensors were not directly programming [9] allows for an efficient use of related. For example, weight does not directly the Bayes rule for conditional probabilities to influence the blood pressure. Instead, we have obtain a diagnosis process from the declarative an estimation of the hydration level of the patient that influences both the weight and the Hence, from a given sequence of observations starting from a known state, we are able to This allows for an even simpler collaboration compute the optimal sequence of states the of the nephrologists with the system: each patient has visited. More precisely, the parameter characterizes only a given state, and Forward algorithm gives the exact probability is only related to a single medical sensor. This of any state at each time step, considering all influence is represented as a graphical curve Once these probabilities have been computed, The first significant difference between the 2 the physician can interpret them directly, since groups is related to the number of visits in each state has an obvious medical meaning. excess from the monthly one. Actually, the test This information is far more precise than the group has paid almost 90% more visits per mere state sequence: it shows the confidence patient than the Diatelic group. This gives an the system gives to each state. For example, a uncertainty factor of 0.66% (ANOVA test). hesitation between two states is clearly visible, Thus, this diminution is very significant. whereas the Viterbi algorithm would have The most important difference, from a medical point of view, is that the Diatelic group has a Alerts are generated from these probabilities in far better controlled blood pressure. In fact, two cases: when the most probable state is not whereas almost all of the patients were slightly the "healthy" one, or if the difference between in hypertension, patients from the Diatelic the two most probable states is negligible. group have decreased their blood pressure by a Therefore, alarms are generated if the system is not sure of the state of a patient, or when an (p<3%). In the same time, they slightly decreased their pills consumption (p<6%). displayed on the patient screen to suggest Even more, the average duration of hospital treatment is almost halved: whereas a patient At the same time, it is added to the medical from the Test group stays for 20 (+/- 36) days team's main screen to draw the physician's at the hospital in a year, a patient from the attention to this anomaly. From this page, the Diatelic group only stays for 11 (+/- 14.5) days medical team can see the alerts from all the at the hospital in a year. Unfortunately, this patients; they can also access all of the difference is not statistically significant, gathered data from the past months, and finally principally because of the large standard adjust the patient's treatment or his profile. 5 The experiment
All these facts amount to an average 14,000 In association with nephrologists from the largely overcome the price of equipping each ALTIR (Lorrain Association for Renal Failure patient with a computer. Unfortunately, the costs of continuous ambulatory peritoneal randomized experiment during 2 years, with 30 voluntary patients spread across 2 groups. treatments. Since these are not significant, The first one (the Test group) is monitored the classical way, while the second one (the Diatelic group) is monitored with the Diatelic Considering all the advantages of the system, system. Each patient was treated by peritoneal we decided to patent it [10] and to create an dialysis for one month at home before his enterprise to continue its development. This integration into the experiment. This way, each year, the experiment will be extended to 300 patient had time to learn the medical procedure patients to study the system applicability to a Gladly, the two groups were statistically 6 Method discussion
homogenous. At the very beginning of the experiment, all the patients were around 70 with partially observable Markov models with years old, with a Charlson index around 5.4. other techniques. However, it is difficult to After two years, 12 patients died, and 6 had quantify exactly the accuracy of a diagnosis. left the experiment for other reasons. There is Therefore, this comparison was based on the no significant difference between the reasons physicians' remarks on the diagnosis accuracy The first method we tested was a standard cannot say if it is suited to the collaboration established in collaboration with the ALTIR nephrologists. The model was very hard to 7 References
1 Jeanpierre, L.: “ Apprentissage et adaptation pour la necessary to implement each rule. The worst modélisation de processus stochastiques reels” , PhD diagnosis was too sensitive to input variations. 2 Rabiner, L.R.: “ A Tutorial on Hidden Markov Models and Selected Applications in Speech To overcome this sensitivity, we upgraded this Recognition”, Proceedings of the IEEE, 1989, 77(2), system with fuzzy logic [12]. In particular, the sensors' model was exactly the same as the one 3 Durand, P-Y., Kessler, M.: “ La dialyse péritonéale we used in the Markov model. Each sensor was based on three fuzzy values depending on 4 Jeanpierre, L., Charpillet, F.: “ Hidden Markov the signal value being normal, insufficient or Models for Medical Diagnosis”, International excessive. Moreover, each rule was given a Workshop on Enterprise Networking and Computing confidence factor based on the importance the in Health Care Industry, HEALTHCOM’ 2002, June physicians gave it into their own diagnosis. However, even this new system was deceptive. 5 Cassandra, A.,. Kaelbling, L.P., Littman, M.: “ Acting optimally in partially observable stochastic In fact, fuzzy logic brought some precision domains” . National Conference on Artificial into the observation process, which was a great enhancement. However, it failed into ensuring 6 Zadeh, L. A.: “ Fuzzy Sets” , Information and a temporal stability of the diagnosis. More precisely, the diagnosis of the past days had 7 Steimann, F.: “ On the use and usefulness of fuzzy disproportionate importance: sometimes the sets in medical” , Artificial Intelligence in Medecine, diagnosis changed totally from one day to another, sometimes it kept the same diagnosis 8 Koenig, S., Simmons, R.G.: “ Unsupervised learning even if the input data had changed a lot. The of probabilistic models for robot navigation”, trouble is this threshold was very sensitive, International Conference on Robotics and and different from one patient to another. A definite trouble of these two expert systems 9 Puterman, M.: “ Markov Decision Processes: discrete was that the medical state of the patient was stochastic dynamic programming” (John Wiley & artificially reported from one day to the next one. More than the insertion of several new 10 Hervy, R., Romary, L., Charpillet, F., Pierrel, J-M, Thomesse, J-P, Petitjean, E., Jeanpierre, L., Durand, thresholds to tune this influence, this way of P-Y, Chanliau, J.: “Système pour le suivi à distance ensuring some temporal stability complicated de patients”, Patents in France (FR2804265, the rules set beyond what we expected. The 07/27/2001), in Europe (WO0154571, 08/02/2001), result was such that no one was able to read and in the U.S.A. (09/539 988, 03/30/2000) Finally, the model parameters were spread all over the rules code, even if we paid particular 12 http://www.iit.nrc.ca/IR_public/fuzzy/fuzzyClips/ attention to regroup most of them. The result Web site of the fuzzy extension of CLIPS. of this was that it was almost impossible to 13 Bellot, D.: “ Fusion de données avec des réseaux bayésiens pour la modélisation des systèmes dynamiques et son application en télémédecine” . During the past year, we compared our results with those obtained with dynamic Bayesian networks [13]. However, even if this model seems to give comparable results in laboratory, no real case use was attempted. Therefore, we

Source: http://laurent.jeanpierre1.free.fr/recherche/papiers/Cimed2003.pdf