In order to properly specify the device intended use an important document to be produced is the General Description, which aims to summarize the main functionalities of the device and to whom the Intended Use refers to. Main points discussed in these documents are reported in the following paragraphs.
Short description of the device / system
Anemia Control Model (ACM) is a software application, which is intended to support doctors in the anemia drug therapy managing. In particular the software is supposed to analyze patient data and to perform a suggestion for the best Erythropoiesis Stimulating Agents (ESA) /Iron pharmacological treatment (dosage and scheduling) for maintaining the hemoglobin level within the target range. Therefore, ACM system provides a dosage/scheduling suggestion for ESA and Iron medicaments. Then doctors are required to evaluate case by case whether the ACM suggestion is safe for the patient; afterwards, basing on their experience, they may formulate the drug prescription following the software recommendation or decide to formulate their own drug prescription.
ACM receives (as an input) essential information from the third party. Then it elaborates the provided information to compute an ESA/Iron therapy strategy. Thus, the ESA/Iron dose suggestion elaborated by the ACM algorithm is sent to the same clinical system as an output. Therefore ACM is not designed to be utilized or administered by an external user, ACM consumer is the third party system. ACM does not store or manipulate any data, only utilizes data to elaborate drug recommendations.
Description of sub-systems / options
ACM is mainly composed of two sub-systems:
- A predictor model which depending on the input data forecasts the response to ESA and Iron therapy for a specific patient;
- An algorithm that using the predictor model extracts the optimal ESA / Iron therapy to achieve the established clinical outcome for anemia management.
List of device composition / materials used
ACM is a computer program wrapped in a service application; it is not part of a physical device, thus:
- No material has been used to manufacture it;
- It does not include any component other that the software itself; - It is not composed by any medicinal substance, tissue or blood product; - Characteristics like sterile condition do not apply in this case;
Development context of the device, Patient population and medical indication
As mentioned in the previous sections, ACM is composed of a predictor model and of a policy extractor. For completeness and to give a proper idea of the kind of information that must be presented in the General Description document, in this section we repeat in a concise way what already discussed in the previous chapters of this thesis.
The predictor model is implemented as a Multi Layer Perceptron. Artificial Neural Networks are computational models inspired by the processing principles found in brains, and extensively used as function approximators in a variety of application scenarios. They are composed of com- putational units called neurons, exchanging information through weighted connections. Neurons are organized in a variable number of layers: an input layer to which the data is fed; an output layer where the result of computation is returned; and one of more hidden layers in between. The ANN, presented with a collection of input-output pairs called the training set, learns by example to approximate the relation between such pairs; to do so, it iteratively adjusts the weights of its connections. If the learning phase is successful, the resulting model will be able to generalize what it has learned to unseen examples: that is, given a new set of inputs (the test set) the ANN should be able to predict the corresponding outputs with reasonable accuracy. In order to ensure good generalization abilities, we applied the typical machine learning paradigm, i.e. data has been divided in three datasets; on for training, one for validation and one for testing. The policy extractor is implemented through an algorithm that consist of:
- A set of environment states S; - A set of actions A;
- A set of rules to reward actions;
- A set of rules to control anomalous situation;
- A selection criteria for the optimal action based on above-mentioned points;
In our case the state S(t) represents patient clinical status at time t, while the action A(t) represent the suggested ESA / Iron dose. The set of actions available to the agent are restricted (e.g., use of available dosages). Basically, given clinical targets provided by experts (or simply implementing guidelines) the policy extractor algorithm simulates different dosages in order to select the one able to achieve these targets.
Specific upper limits of ESA and Iron dosages, as well as hemoglobin upper limits for ESA therapy interruption have been imposed to the machine as specified by the anemia therapy guidelines, so that dangerously high drug doses, are automatically blocked and no ESA therapy
comorbidities, gender or age, etc.) does not influence the inclusion of the patients. ACM has been created to facilitate the prediction of a correct anemia therapy with the lowest error. Therefore, although all HD patients are generally included in the study, the ACM system finds a fundamental application on a special HD population, namely patients suffering Chronic Kid- ney Disease (CKD) secondary anemia (defined as low levels of plasma hemoglobin) receiving a pharmacological treatment for correction (i.e., ESA therapy and/or Iron supplementation). Moreover, special conditions may occur and some patients may become temporary not eligible for ACM prediction. They will not be excluded from the study; however, ACM analysis may be suspended, case-by-case, until all the necessary criteria are satisfied again. For any HD patient in the clinics the secondary inclusion criteria are:
- A history of dialysis therapy >= 3 months;
- Enough clinical and biochemical records during the 3 months preceding the ACM elabora- tion;
- No blood transfusion during the 3 months preceding the ACM elaboration;
- Unique administration of intravenous (IV) Darbepoetin alpha as ESA during the 3 months preceding the ACM elaboration;
- Unique administration of intravenous (IV) Iron preparation during the 3 months preceding the ACM elaboration;
- Absence of errors in the input data format;
Although any ESRD patient who satisfies the above-mentioned conditions may be ideally in- cluded in the present study, for safety reasons, we are going to validate the ACM system on a restricted population. Hence, in a first phase of the evaluation, only patients from four pilots clinics (one in Portugal; two in Spain; one in Czech Republic) will be initially included. There- after, depending on the study outcome, we will be able to extend the use of ACM on the general ESRD population.
What is intended use of the device
The following ACM intended uses are the closest to the definition of a medical device:
- Recommendation of the best ESA/Iron drug dosage and scheduling; - Anemia therapy outcome prediction.
Major aim of ACM is maintenance of patient hemoglobin level within established targets assuring patients’ safety; hence, its utilization respects the following policy:
- ACM suggestions cannot become actual prescriptions without external medical interven- tion;
- ACM utilization implies continuous clinician supervision and control action so that expert nephrologists always perform prescriptions after viewing both the software outcome and the patient clinical conditions;
- ACM algorithm has been trained on real patient retrospective data, i.e. actual biochemi- cal/clinical data together with actual administered drug quantities. Therefore, the system has learnt true therapeutic policies and it elaborates new data basing on that past experi- ence;
- Specific lower and upper limits ESA or Iron dosages as well as hemoglobin upper limits for ESA therapy interruption have been imposed to the machine as specified by the anemia international guidelines, so that dangerously high drug doses, are automatically blocked and no ESA therapy is administered when hemoglobin concentration is over 13 g/dl; - It must also be mentioned that, since ACM has been trained on real doctor prescriptions
it is programmed for respecting the limits learnt from the real medical experience. Indeed, during the preliminary validation phase, ACM never suggested a dose out of the normal range used by doctors.