I thoroughly enjoyed the paper by Drs. Gilberto Câmara and Antonio Miguel Viera Monteiro and hope that more researchers will be enticed by the main ideas presented above. I hope to see a stronger cross-fertilization of this emerging interdisciplinary field, connecting the use of so-called intelligent systems to spatial health data analysis.

The difficult task is to sum up and provide a brief discussion of this paper. Reading the first part I learned a term with which I had little or no familiarity - geocomputation - posited by the authors as a new interdisciplinary field using computer-intensive methods, including neural networks, fuzzy logic, genetic algorithms, and cellular automata for spatial data analysis.

The study of spatial and spatial-temporal epidemiological data is a timely issue which is driven by both decreasing technology costs and increasing availability of information. For example, it is becoming increasingly possible to access georeferenced public health data in a speedy manner through the Internet for analyzing and merging with other information. Several models and methods to work with spatial health-related data have cropped up in the literature in the last twenty years. Most of these were developed in other areas, like geostatistics, which originated in the mining industry and was later borrowed to help understand and explain the spatial distribution of health events. As is common in many applied sciences, the method is first introduced in an intuitive way, and once the heuristic results prove encouraging, there is major involvement by mathematical and statistical theorists to get the technique soundly established. The wave of progress following this pattern continues with Câmara and Monteiro's paper, presenting a basic review of existing possibilities for the use of different computing procedures to perform spatial health data analysis.

On the application side, I would partially support the motivating statement of the paper citing Oppenshaw (1996), that "many end users merely want answers to fairly abstract questions ...". However, some care should be exercised here. Some twenty years ago I heard in a Brazilian workshop on statistical methods for epidemiologists, particularly on multiple regression, that the basic concepts are cumbersome and difficult to be understood by public health workers, and that they should be more involved in collecting good data to be analyzed by the "foreigners", i.e., specialists in statistics. Obviously, the authors of the paper would not wish us to merely engage in using these "black box" tools (which are well understood in the artificial intelligence community) but rather, that we begin close collaboration to both further the knowledge of these new methods and convince ourselves that they could be included in the analytical tool box of epidemiologists and public health professionals.

The authors provide examples of real analyses in the hope of giving a genuine applied flavor to the methods reviewed. I wish to make some comments on these applications. The first concerns the use of the GAM (Geographical Analysis Machine) to find clusters in data that are originally areal data. Although the authors emphasized that it is only an example, there is no mention of the large differences in area sizes and population distribution in Rio de Janeiro's districts, which I believe could substantially influence the results. If one uses some sort of altered or transformed data set, one must interpret it with caution and be certain that the alteration is stated clearly to avoid misuse by newcomers to this field of spatial analysis research.

My other point concerns Section 3, on neural networks and geographical analysis, where the authors present a classification problem to produce a map of environmental vulnerability. One of the most fundamental aspects of neural network modeling is the requirement of "plenty of training data", which is properly identified in the paper. Neural networks are "adaptive computing" in that they learn from data to build a model. Therefore, the training data set should contain all examples of possible sets of explanatory and outcome variables if one uses the workhorse of neural network modeling: a feed forward network with a back propagation algorithm. Users interested in applying this new technology should be aware of this important aspect. In addition, analysts must be willing to both tolerate the large amount of time for training and have a "black box" model which unfortunately does not provide the ability to explain the reasoning used to arrive at a result. This still limits the usefulness of this technique in some areas, particularly when one is interested in measuring the effects of input variables rather than prediction.

Recent developments in computing performance have provided a wealth of opportunities for advancement of new analytical approaches to spatial data analysis. These include the increasing use of Bayesian thinking, particularly with the introduction of the Monte Carlo Markov Chain (MCMC) approach to tackle intractable integrals. For the unfamiliar reader, the paper provides a brief introduction to various techniques. Some of these techniques were derived from the so-called intelligent systems, and it is the hope of the authors, and also mine, that they may assist our capability to convert data into information.