Machine learning is an interdisciplinary field combining computer science and mathematics to develop models with the intent of delivering maximal predictive accuracy. This is done by detecting patterns from an incomplete set of examples composed of past data. For this reason, this is a data-driven discovery process. The quality of the predictions of machine learning algorithms rely mainly on the number of samples and the quality (amount of information contained) of the variables used to describe each sample of the phenomena (Bishop 2007). The following books are good references of these techniques and their applications (Bishop 2007, Kelleher et al. 2015, Schölkopf and Smola 2001, Vapnik 1998). Machine Learning is very promising in the field of neuroimage, where techniques as artificial neural networks, support vector machine, random forest, k-means, nearest neighbor, decision tree have been successfully applied (Huang and Chaovalitwongse 2015, Wernick et al. 2010). Currently Deep Learning Techniques have attracted the attention of the scientific comunity (Chen and Lin 2014). In this project we are going to explore both classical and novel machine learning techniques as a data scientist to find the best suitable techniques in a data mining process.

With medical data, there exist several practical, technical and ethical issues with acquiring great amounts of examples compared to other research fields such as information retrieval where millions of examples are freely available. Therefore, it is of utmost importance to extract the maximum amount of information of each example. This is why an exhaustive feature extraction phase is done.

Several machine learning techniques have been used in medicine in the past to diagnose patients (Hu et al. 2013), to provide support to the radiologist in the detection of aneurysms (Suniaga et al. 2012) and to predict radiation therapy outcomes among other things. Pubmed indicates an 52.4% increment in the number of publications (from 1204 to 1911) related with the keywords “machine learning” between years 2014 and 2015, being this a very active field of research.

Depending on the problem, different kinds of features are needed. Nowadays, machine learned features extracted from raw data are becoming more used, as deep learning approaches have been very successful. However, the interpretation of those features is an open problem, as the clear understanding of the role of each feature is what could give a deeper insight into the aneurysms causes and risk factors.

Up to now, what has been used in features extraction methods based on angiographic image analysis is the following: the first step in the feature extraction step corresponds to the vascular geometry creation. Using the angiographic images a three-dimensional triangulated surface is obtained using an automatic segmentation method based on geodesic active regions in combination with an image standardization technique (Hernandez and Frangi 2007). Employing this model, different features are extracted. Morphological features, such as the aneurysm aspect ratio, non-sphericity index (Raghavan et al. 2005), volume and surface. Other more sophisticated descriptors to characterize its shape are the three-dimensional Zernike moment invariants (Millán et al. 2007). Hemodynamics (blood flow) is also analyzed extracting information like maximum and mean velocities defined at peak systole and averaged over the cardiac cycle, this based on a blood flow simulation that use the three-dimensional geometry obtained in the first step (Villa-Uriol et al. 2011). Other measurements computed on the model surface are areas of elevated pressure, impingement jet location, maximum static pressure and the wall shear stress (Villa-Uriol et al. 2011). A problem is that these measures depend on the location of the aneurysm neck, which is not detected reliably by the existing automated algorithms (Villa-Uriol et al. 2011). Additionally, an accurate structural analysis of the blood vessels would possibly have the best predictive capabilities. Nevertheless, in practice it requires too many assumptions and simplifications, not being sufficiently reliable (Ma et al. 2007).

A possible contribution to the state of the art would be a better aneurysm neck detection method. Also, it has been noted in (Villa-Uriol et al. 2011) that the feature extraction process could be enhanced by the use of additional information of the patient, such as known diseases that could give prior information to take into account, instead of using the same model for everyone. A different approach is to optimize different existing methods, as most of them are time expensive, taking several minutes, being sub-optimal for the very high loads in hospitals.

Having the features to feed a machine learning algorithm, the predicted rupture risk can be computed. However, to understand the underlying principles involved in the rupture process, having a great amount of features is counterproductive, since most features may not be involved in the process. The process of finding the most relevant features is known as feature selection.

Nowadays, the reasons of aneurysm formation and rupture are unknown. Therefore, even if the feature selection process may only give a set of important features (in a statistical sense) without any further descriptions, it may give some hints about what variables are correlated with the rupture in dependent (if they only affect the process if other variables have a range of values) and independent ways.

Bisbal et al. (2011) initiated work on cerebral aneurysm description using data mining approach based on demographic, clinical and images data on 157 cases. However, there are several aspects to improve in the feature selection process, such as not assuming independency and linearity between the different variables. For example, in (Izbicki et al. 2011) variable pairs where found to be significant for assessing the rupture risk of abdominal aneurysms. It is also important to take the different error types into account in order to define with greater precision the performance of the proposed prediction.

Aneurysm causes - state of the art (March 2017 – May 2017)

At the beginning, the first task is to confirm that our prior conception of the information that is hypothesized to be relevant to estimate the risk of rupture is exhaustive. This information will be used to guide the data collection and feature extraction processes.

Hospital data collection (March 2017 – June 2017 retrospective, December 2018, December 2019 prospective)

The first step will be focus on data recollection from the Hospital Carlos van Buren (HCvB). It is important to mention that this project is inserted in a work team, namely in the HCvB: Dr. Pablo Cox and Dr. Rodrigo Riveros, interventional neuro-radiologists, and RT Maximiliano Godoy. HCvB is the public hospital of the Valparaíso region and this Hospital is a center of reference in neurology and neurosurgery in Chile.

The proposal is to undertake a retrospective and prospective study, to include patients from 2014 to 2016 retrospectively and from June 2017 up to December 2019 prospectively, counting on the approval by the local Ethical Committee. Inclusion criteria will be based on presenting a confirmed diagnostic of cerebral aneurysm. Patients will be enrolled through the Angiography Department of the Hospital, by the interventional neuro-radiologists of this project. No change will be applied in the way the patient is diagnosed or treated, only his/her data will be included in the present study. According to the registries from previous years, the Angiography Department of HCvB receives 100 patients per year with diagnostic of cerebral aneurysm, 83% of them ruptured. Specifically from 2014 to 2015, 238 patients with aneurysms were diagnosed in Angiography of the HCvB. In the prospective period of 2.5 years, we are expecting the additional inclusion of 250 patients.

Data to be collected are

• Angiographic images (already in DICOM format)

• Informs: radiologist inform to establish diagnostic; radiologist inform of intervention (already in

  digital format)

• Health records: these data are already accessible in a digital home-made database in the Angiography Deparment of HCvB, based on FileMaker, developed and administrated by RT Maximiliano Godoy

  • Patient demographic information. Age, gender, weight, height, BMI, city of residence (to be related to epidemic data and weather data)

  • Diagnostic (ICD10) and comorbidity

  • Clinical and treatment information: Glasgow score, drugs used (which ones in which concentrations, or which suspended), allergies, number of punctures, etc.

  • Laboratory analysis, in particular PCR, among others

• Epidemic, from the Ministry of Health (Departamento de Estadísticas e Información de Salud DEIS http://www.deis.cl) : in particular, registry of seasonal variation (syncytial virus, and others).

• Weather data, in combination with the epidemic information. The “Centro de Ciencia del Clima y la Resiliencia” (http://www.cr2.cl) has released in 2016 a tool to access the Chilean historical weather data, including variables, such as mean, maximum and minimum temperatures and precipitations across the country, including at least 229 weather stations. In this step it is necessary to obtain all the information available since year 2014, so it can be later used jointly with the aneurysms rupture dates.

The inclusion of epidemic and weather data is motivated by the recent observations of neuroradiologist that there seems to be a peak of aneurysm ruptures observed in relation with seasonal fluctuations (syncytial virus for instance). The underlying hypothesis is that rupture might be influenced by mechanisms of inflammation, in combination with hemodynamic and biomechanical stresses exerted over the arterial wall.

Data storage (May 2017 – July 2017)

The Proyecto de Mejoramiento Institucional UVA1402 for the Development of an Interdisciplinary Platform for Healthcare Innovation being executed by the institution will provide a secure database to store the information. The objective of this task is to centralize the access of all the data in a unique database and to anonymize the data, so only the relevant clinical professionals may access the identity of the person behind specific data.

The initial computational framework implementation (client-side) will also be considered for the retrieval and the processing of the data.

Technical report (May 2017 – August 2017)

Write a report to summarize the aneurysm rupture information related to each resource and a detailed data description from each source for future reference.

Medical images feature extraction (May 2017 – April 2019)

The initial problems detected in the feature extraction stage of the image characterization problem will be further explored, such as robust aneurysm neck detection, fast and robust three-dimensional mesh computation and new feature engineering to describe the aneurysm, since features like the non-sphericity index have been shown to have a significant correlation with the aneurysm rupture (Villa-Uriol et al. 2011). A new promising feature that will be pursued is the automatic estimation of the vascular age of the patient, as the two experienced interventional neuroradiologists members of this project are able to qualify the patient’s vascular age as a function of the artery tree, visible on the angiography.

The first step, after briefly exploring the existing problems and their solutions, will be the initial build of a flexible and modular computational framework for image processing, capable of easily integrating external libraries to use whenever is possible existing software solutions, such as ITK (Insight Segmentation and Registration Toolkit) and VTK (The Visualization Toolkit), to avoid re-implementing algorithms, since it is a time demanding task. Naturally, the integration of different external solutions will be done as needed, but it is important to lay a robust foundation to allow this to happen seamlessly.

Considering that, the methodology steps will be applied to each of the following tasks:

• Initial computational framework implementation for image processing

• Robust aneurysm neck detection

• Vascular tree description

• Fast and robust three-dimensional mesh computation

• Morphological aneurysm description

• Additional feature consideration

Additionally, a first paper will be written to be published in a conference of level A or A/B in the CORE system or in a journal (April 2019).

Other data feature extraction (July 2018 – April 2019)

The rest of the information is composed of a plethora of different features, such as time series (weather and epidemic data), categorical (smoking habit) and numerical information (age, height). To take advantage of this diversity, it is important to carefully extract relevant features, so they provide additional information.

The following tasks are identified for which the proposed methodology will be applied:

• Relevant features for weather data extraction, such as humidity, temperature and their changes in short time intervals in specific locations

• Epidemic data feature extraction

• Health records information extraction

• Dimension reduction by finding correlated variables

Feature combination for aneurysm description (May 2019 - October 2019)

Using all the extracted features, the aneurysm rupture risk will be modeled using a data-driven approach. It will be very important to take into account the different kind of features and missing values, as it has been shown that in the classifying process this can have a very significant impact (Bisbal et al. 2011).

Specifically, the following task have been identified:

• Missing value estimation (November 2018 – January 2018)

• Different feature space combination (February 2018 – April 2018)

Aneurysm rupture risk estimation according to combined features (May 2019 – October 2019)

To make a good estimation, it is important to use or modify a classifier that is able to take into account the risks of each decision and give a confidence on its estimation.

• Imbalanced classes impact quantification and consideration (May 2019 – June 2019)

• Error cost inclusion (June 2019 – July 2019)

• Classifier confidence estimation (July 2019 – August 2019)

• Rupture risk prediction (August 2019 – September 2019)

The publication of the experiments results and insights acquired will be published in an ISI indexed Journal of the area article (April 2019 – September 2019).

Feature filtering and selection to identify the quantitative role of the individual features (September 2019 – February 2020)

After obtaining a good classification accuracy compared to the existing literature, the objective is to identify the subset of the most relevant features and their correlations that could explain the rupture of cerebral aneurysms. The tasks to achieve this objective are:

• Find lineal and non-lineal correlation among different variables sets and the predicted aneurysm rupture risk. This should take into account variables of the same source and multiple sources (September 2019 – November 2019)

• Decrease the number of used features to a minimum maintaining the classifier’s predictive precision (December 2019 – January 2020)

The publication of the experiments results and insights acquired will be published in an ISI indexed Journal of the area article (October 2019 – February 2020)

The sponsoring institution is executing the PMI (Plan de Mejoramiento Institucional) UVA1402 for the Development of an Interdisciplinary Platform for Healthcare Innovation . The actual available equipment includes two multi-processor servers with more than sufficient storage for all the required data of the study and services to store sensible data. Additionally, personal powerful workstations are available for heavy processing tasks. In this context, the data are stored within the custody of Universidad de Valparaiso’s Informatics Division (DISICO), with physical and digital safeguard.

Physical space, access to the university library, internet connection and relevant journals subscriptions are also available.