Antonina Starita’s research while affiliated with University of Pisa and other places

3945 Poster Board III-881 Introduction Non Hodgkin Lymphoma could be clinically divided as low grade/indolent NHL (LG NHL) and high grade/aggressive NHL (HG NHL). These diseases are chemo and radio-sensitive and improvements have been achieved by immunotherapeutic approaches. However some patients will relapse and a follow up strategy has to be planned in order to detect and treat them. Several aspects should be considered in planning a follow up including safety, specificity, sensitivity, costs and impact on the patient's psychology: an optimal follow up should mediate between these ones. The more diffuse follow up have been planned years before the introduction of innovative methods and imaging techniques, suggesting the opportunity to revise these programs. Methods We collected data about 418 NHL patients -both low and high grade- treated at our institution from 1990 to 2005 who achieved a complete remission according to Cheson criteria and who entered a follow up program which schedule is planned for 5 years divided in two periods: first two years evaluation every 3 months and in the following three years every sixth month. At each visit physical examinations, blood testing (blood count, chemistry) are performed; for imaging techniques we alternate a whole body CT scans to ultrasounds and chest X-ray coupled. Analyzing time to relapse (TTR), we tried to optimize our follow up schedule trough a computation application known as multi-objective analysis. The first step of this method has been to choose and try to quantify the costs of a follow up which reflect its effectiveness. We considered as costs the expected time between relapse and its detection (Ca) and the expected number of performed examinations before failure or censoring occurs (Cb). The total follow up costs could be summarized in a vector Cref: (Ca,Cb). After doing that we described survival analysis, relapse rate and their onset time in order to be suitable for the informatic analysis. We used a log logistic parametric model to do that. Next we shaped our ideal follow up as a structure based one, which is currently the most used in medical practice: a first period where examinations are more tightly spaced, followed by a second period where they are performed further apart. We describe such follow-up schedule “S” as (d1, k, d2) : d1 is the time between the first k examinations; d2 is the time between the remaining examinations. Applying the log-logistic model we calculated the Cref and schedule for our follow up: Cref was (Ca,Cb)=(8.5,6.2); this means that, on average, every patient entering the follow-up will perform 6.2 examinations, and among those who incur in a relapse, the relapse will be detected 8.5 units of time (eg. weeks) after its onset. The follow up structure was (13,8,26). We then calculated all the possible combination of (d1, k, d2) values from 1 to d1 = 25, k= 24 and d2 = 60. Results 360000 follow up schedules had been detected after searching all the possible combinations for d1, k and d2. For each one Ca e Cb costs have been calculated and than compared by multi-objective analysis to those of the current schedule, We look for both follow up structured and “free” follow up, where “free” means that intervals between examination is continuously variable. When comparing follow-up schedules, we apply the rule of Pareto-dominance: the schedule S1 is superior to the S2 schedule if and only if the cost values for S1 are both lower than the cost values for S2. The method then takes into account only those schedules which improve the current one, with respect to both the Ca and the Cb costs. After multi-objective analysis six follow up were detected. These were as following: (16,10,22), (15,8,20), (15,6,19), (16,8,19), (16,5,18), (16,4,18) and are shown in Fig. 1. Conclusions no differences in follow up schedules emerged when we considered separately LG and HG lymphoma patients. Maximum improving of our follow was just 4% and all the new schedules had wide time frequency visit in the first period and a narrow one in the second period: this was a consequence of the higher relapse distribution in the first period and this follow up organization lead to a lower total number of visit without risk of lose any relapse. This is the first application of this analysis to hematological patients confirming the validity of a follow up schedule should be shaped in two different period. Disclosures No relevant conflicts of interest to declare.

4772 Introduction Non Hodgkin Lymphomas (NHL) are commonly divided in two large groups known as high grade (HG) and low grade (LG) lymphomas. Specific therapeutic strategies are potentially curative, but a tailored treatment should be planned from the diagnosis on the basis of the prognostic features. Some clinical features at the diagnosis or relapse may be grouped to build prognostic indexes. These indexes are “build” starting from features present at diagnosis and related to patients and disease onset, but lacks to give any information about disease history and chemosensitivity. For these purposes others methods have been evaluated and probably the two that gave major help to detect slow responders or resistant patients are minimal residual disease (MRD) and PET scans. Here we develop and apply a partitioning recursive algorithm, known as HCS, by which possibly detect new possible prognostic factors. Methods Our dataset comprised 651 NHL patients followed at our Institution from 1990 to 2005, divided in High grade (HG NHL, n=343; 52,7%) and Low grade Lymphoma (LG NHL, n= 308; 47,3%). Only patients who enter in a follow up program were considered: therefore patients with at least a partial response. We considered as variables for algorithm analysis: age, sex, histological subtype, IPI status and bone marrow involvement, treatment approaches (poli-chemotherapy, mono-chemoterapy, radiotherapy, surgery, purine based chemotherapy, monoclonal antibodies, transplant approach, oral chemotherapy), response to therapy, previous successful treatments, previous relapses, previous failed therapies. Data were analyzed by a recursive partitioning algorithms. A partitioning recursive algorithm. This tool is thought to splits data in different subgroups that behave in a different way. It works starting from data and utilized them to create all the possible combinations of splits available. Among them it chooses the best one by statistics and finally applies it to patients' datasets. For these reason it is defined as “partitioning”. Afterwards the algorithm starts again the analysis on the subset previously detected and that's because it is called “recursive”. Results The most important split emerged to be the quality of response: patiets were splitted between patients in partial remission (PR) and complete remission (CR). Among PR patients, one subset (subset 1) with worse prognosis were found: both comprised patients with HG NHL not treated with monoclonal antibodies and/or transplant. Therefore the remaining PRs patients, treated by transplant approach and immunotherapy, had a better outcome. In CR group subset 2 has been detected, comprising HD NHL patients not treated by oral chemo alone were detected. Those patients, who had a better outcome, had been treated aggressively with polichemotherapy and/or autologous transplant. Differences between detected groups are statistically significant with p. value 0,03 maximum. Splits are shown in Figure 1 and 2. Conclusion Application of computer science analysis to NHL patients has been successfull. Quality of response emerged as the most important prognostic factor but among both PRs and CRs patients those with better outcome had HG NHL diagnosis and were treated by autologous transplantation or/and immunotherapy. This analysis confirms data available about transplant as a good approach for NHL patients. Disclosures No relevant conflicts of interest to declare.

We propose a new method based on a Recursive Neural Network (RecNN) for predicting polymer properties from their structured molecular representations. RecNN allows for a completely novel approach to QSPR analysis by direct adaptive processing of molecular graphs. This model joins the representational power of structured domains with Neural Network ability to capture underlying complex relationships in the data by a process of training from examples. To this aim, a structured representation was designed for the modelling of polymer structures. The adopted representation can account also for average macromolecule characteristics, such as degree of polymerization, stereoregularity, comonomer distribution. To begin with, this model was applied to the prediction of the glass transition temperature of (meth)acrylic polymers with different degree of main chain tacticity. The results so far obtained indicate that the proposed representation of polymer structure can convey information on both the repeating unit structure and average polymer features. The ability of the proposed RecNN method of treating this structured representation makes this method more general and flexible with respect to standard literature methods. Moreover, the same model can handle at the same time the Tg of polymer samples present in only one tacticity form together with that of polymer with different stereoregularity.

We report here a recent study on the prediction by recursive neural network of the glass transition temperature of (meth)acrylic copolymers, for which appropriate structured representations are proposed. It is shown that the flexibility of such description allows for simultaneously treating different classes of compounds as well as accounting for different average properties such as tacticity and molar composition.

This paper reports some recent results from the empirical evaluation of different types of structured molecular representations used in QSPR analysis through a recursive neural network (RNN) model, which allows for their direct use without the need for measuring or computing molecular descriptors. This RNN methodology has been applied to the prediction of the properties of small molecules and polymers. In particular, three different descriptions of cyclic moieties, namely group, template and cyclebreak have been proposed. The effectiveness of the proposed method in dealing with different representations of chemical structures, either specifically designed or of more general use, has been demonstrated by its application to data sets encompassing various types of cyclic structures. For each class of experiments a test set with data that were not used for the development of the model was used for validation, and the comparisons have been based on the test results. The reported results highlight the flexibility of the RNN in directly treating different classes of structured input data without using input descriptors.

In this paper we present a necessary and sufficient condition for global optimality of unsupervised Learning Vector Quantization (LVQ) in kernel space. In particular, we generalize the results presented for expansive and competitive learning for vector quantization in Euclidean space, to the general case of a kernel-based distance metric. Based on this result, we present a novel kernel LVQ algorithm with an update rule consisting of two terms: the former regulates the force of attraction between the synaptic weight vectors and the inputs; the latter, regulates the repulsion between the weights and the center of gravity of the dataset. We show how this algorithm pursues global optimality of the quantization error by means of the repulsion mechanism. Simulation results are provided to show the performance of the model on common image quantization tasks: in particular, the algorithm is shown to have a superior performance with respect to recently published quantization models such as Enhanced LBG [Patané, G., Russo, M., 2001. The enhanced LBG algorithm. Neural Networks 14 (9), 1219–1237] and Adaptive Incremental LBG [Shen, F., Hasegawa, O., 2006. An adaptive incremental LBG for vector quantization. Neural Networks 19 (5), 694–704].

Determining a compact neural coding for a set of input stimuli is an issue that encompasses several biological memory mechanisms as well as various artificial neural network models. In particular, establishing the optimal network structure is still an open problem when dealing with unsupervised learning models. In this paper, we introduce a novel learning algorithm, named competitive repetition-suppression (CoRe) learning, inspired by a cortical memory mechanism called repetition suppression (RS). We show how such a mechanism is used, at various levels of the cerebral cortex, to generate compact neural representations of the visual stimuli. From the general CoRe learning model, we derive a clustering algorithm, named CoRe clustering, that can automatically estimate the unknown cluster number from the data without using a priori information concerning the input distribution. We illustrate how CoRe clustering, besides its biological plausibility, posses strong theoretical properties in terms of robustness to noise and outliers, and we provide an error function describing CoRe learning dynamics. Such a description is used to analyze CoRe relationships with the state-of-the art clustering models and to highlight CoRe similitude with rival penalized competitive learning (RPCL), showing how CoRe extends such a model by strengthening the rival penalization estimation by means of loss functions from robust statistics.

A novel neural network clustering algorithm, CoRe, is benchmarked against previously published results on a breast cancer data set and applying the method of Partition Around Medoids (PAM). The data serve to compare the samples partitions obtained with the neural network, PAM and model-based algorithms, namely Gaussian Mixture Model (GMM), Variational Bayesian Gaussian Mixture (VBG) and Variational Bayesian Mixtures with Splitting (VBS). It is found that CoRe, on the one hand, agrees with the previously published partitions; on the other hand, it supports the existence of a supplementary cluster that we hypothesize to be an additional tumor subgroup with respect to those previously identified by PAM.

Seizure-free EEG signals recorded from epileptic children were compared with EEG signals recorded from normal children. The comparison was based on the detection of transient events characterized by decrease in the correlation between different traces. For this purpose, a conceptually and mathematically simple method was applied. Two clear and remarkable phenomena, able to quantitatively discriminate between the two groups of subjects, were evidenced, with high statistical significance. In fact, it was observed that: (a) The number of events for the epileptic group was larger; (b) Applying restrictive criteria for event definition, the number of subjects in the epileptic group presenting events was larger. The results support the hypothesis of a decrease in brain correlation in children with epilepsy under treatment. This confirms the efficacy of the EEG signal in evaluating cortical functional differences not visible by visual inspection, independently of the cause (epilepsy or drugs), and demonstrate the specific effectiveness of the analysis method applied.

HER2-positive breast cancer is characterized by aggressive growth and poor prognosis. Women with metastatic breast cancer with over-expression of HER2 protein or excessive presence of HER2 gene copies are potential candidates for Herceptin (Trastuzumab) targeted treatment that binds to HER2 receptors on tumor cells and inhibits tumor cell growth. Fluorescence in situ hybridization (FISH) is one of the most widely used methods to determine HER2 status. Typically, evaluation of FISH images involves manual counting of FISH signals in multiple images, a time consuming and error prone procedure. Recently, we developed novel software for the automated evaluation of FISH images and, in this study, we present the first testing of this software on images from two separate research clinics. To our knowledge, this is the first concurrent evaluation of any FISH image analysis software in two different clinics. The evaluation shows that the developed FISH image analysis software can accelerate evaluation of HER2 status in most breast cancer cases.