Data

Tools

Indexes

Applications

Experiments

Awesome

An open API service indexing awesome lists of open source software.

https://github.com/blackcub3s/msc-finalthesis

The most important programming files, code functions and data processing pipelines for the Machine learning final thesis of my Master's degree. Also, the LaTeX code of the thesis.
https://github.com/blackcub3s/msc-finalthesis

data-analysis latex machine-learning numpy python sklearn

Last synced: about 1 month ago
JSON representation

The most important programming files, code functions and data processing pipelines for the Machine learning final thesis of my Master's degree. Also, the LaTeX code of the thesis.

Awesome Lists containing this project

README 

          
# Description



This repository contains my MSc's final thesis, a quick explanation of its results, and some of the most important Python scripts that I made towards the completion of the thesis.



# Extended Summary



My final thesis “Development of a single-Subject Predictive model of Alzheimer's disease using fMRI and machine learning techniques in individuals with Mild Cognitive Impairment” was an endeavour that involved working with big data, fMRI and machine learning models. I used Python to carry out the data analysis from data retreived from  a neuroimaging project called ADNI[^8].



For the data analysis I took a multi processing pipeline approach, using a leave one-out cross-validation procedure (the data was scarce so there was a need to reduce dimensionality). In order to select the best fitting algorithm, there were several processing pipelines, and for each one of them a different machine learning model was fitted: artificial neural network (multilayer perceptron), logistic regression, naive bayes, random trees, etc.



The conclusions of the master thesis were that Alzheimer's disease has

potential to be predicted by using fMRI in a population at risk for developing it (mild

cognitive impairment) with a decent accuracy. This can be done using machine learning

and an abstract computational neuroscience concept called functional connectivity (an

indirect measure on how interconnected a brain is). However, further research should

be done and results are to be interpreted with a grain of salt, given that although best

fitting model accuracy was good and specificity was high, the sensitivity of the

classification was low, probably due to the low sample size of mild cognitive impairment

patients who would turn into Alzheimer’s disease as opposed to the sample size of

those who wouldn’t.



- Tech stack: Python 3. Used modules: sklearn (machine learning), matplotlib

(data visualization), seaborn (enhanced layer for matplotlib), numpy (matrix laboratory),

pandas (data science). In order to write my final thesis I used LaTeX as, to the best of my

knowledge, is the highest typographic quality system available as open source.



# Abstract (for a more technical read)



In the last years scientists have tried to develop predictive models for

forecasting a future onset of Alzheimer's disease (AD) in people with Mild

Cognitive Impairment (MCI), using fluid, imaging and neuropsychological

biomarkers. To the best of our knowledge, the question of whether functional

magnetic resonance imaging (fMRI) can serve as a viable predictor for

the aforementioned disease in MCI individuals remains unanswered. We have

developed a single-subject predictive model using, for each individual, solely

a rsfMRI scan obtained at the screening visit of the ADNI and follow-up

longitudinal data about future outcomes (AD presence/absence). We included

in our model 23 MCI-c (mean time until conversion: 1.65 years) and

51 MCI-nc patients (mean time of follow-up: 4.61 years). Scan preprocessing

pipelines consisted on registering each scan to Shen's Atlas (214 ROIs),

extract functional connectivity measures and use them as predictors (either

with or without dimensionality reduction) and pair them with the future

outcomes to train and test supervised machine learning models via 10-fold

cross-validation. We found conversion from MCI to AD can be predicted

from rsfMRI with a multilayer perceptron with no dimensionality reduction

with a reasonable accuracy (77.03%, 95% CI from 67 to 87%), good ROC

AUC (0.81), very high specficity (specMLP = 90.20, 95% CI from 83 to 97%)

but weak sensitivity (47.83%, 95% CI from 36 to 59%). Logistic regression

and the Support Vector machines also obtained reasonable diagnostic accuracies

(75.68%, both of them). The models cannot be deployed to clinical

practice yet: further research is needed to increase sensitivity.



# Folders and files to comment



The most important files and folders the reader can refer to, to see either the full thesis or the code with which I got the results from the reader can be headed to any or all of the following folders or files (see thorough explanation for each relevant file):



- [TFM_FINAL_santiagosanchez.pdf](https://github.com/blackcub3s/MSc-FinalThesis/blob/main/TFM_FINAL_santiagosanchez.pdf): the final thesis as was handed in to the university.

- [Writing](https://github.com/blackcub3s/MSc-FinalThesis/tree/main/Writing): Contains the LaTeX files and figures that were put together in order to compile * *TFM_FINAL_santiagosanchez.pdf* *

- [DataAnalysis](https://github.com/blackcub3s/MSc-FinalThesis/tree/main/DataAnalysis): this folder contains the most important programs I made in Python to do the analysis. Within the sub folder [inContext](https://github.com/blackcub3s/MSc-FinalThesis/tree/main/DataAnalysis/inContext) you'll find the same programs, but in context with some of the raw data behind (and other files that contributed to the development of the final thesis)[^1]. The most important files are linked here and you can look into them:



    * [mira_dades.py](https://github.com/blackcub3s/MSc-FinalThesis/blob/main/DataAnalysis/inContext/mira_dades.py): In this script the most important functions are these:

        1. Merge two datasets with complementary information: you can see `IMPORTANT_depura_ADNI_MERGE_i_ARXIU_fMRIs()`

        2. See which subjects convert to alzheimer's and which do not: see `temps_de_conversio(SUBMODALITAT_fMRI)` 

    * [a. fes_merge_en_tensor_3d (obte l'array 3d).py](https://github.com/blackcub3s/MSc-FinalThesis/blob/main/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/a.%20fes_merge_en_tensor_3d%20(obteArray3d).py): In this file the most important snippet of code is the funcion `fes_merge(n_timeseries)` (see down below). 

        https://github.com/blackcub3s/MSc-FinalThesis/blob/bd8e13d5ad14fc7103469cd7c6f38da0e5008288/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/a.%20fes_merge_en_tensor_3d%20(obteArray3d).py#L39-L114



        On the one hand, this function went to each folder name of syntax *"SUBJECTID__UID"* and got the *"SUBJECTID"* (identifier for subject) and *"LLUID"* (identifier for scanner type made to a given subject) to save them in list objects and finally store them in respective .txt files (`__subjectes.txt` and `__uids.txt`). On the other hand, the function also went into each subject folder to find `total.txt` a file that contained fMRI data extracted from .NIFTI files *after* registering that information to Shen's atlas[^3] via FSL. The information inside each subject's `total.txt` was structured as shape `(214,140)` (the first dimension being the Regions of interest (ROIs[^4]) of Shen's atlas -214- and the second one was the number of time series data points available -140 time series data points for each ROI-) and finally I stacked together all these matrices as a numpy ndarray object (*.npy*) to have it all within one single file ([__arr_ADNI_3d_preprocessada.npy](https://github.com/blackcub3s/MSc-FinalThesis/blob/main/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/__arr_ADNI_3d_preprocessada.npy)) of shape `(93,214,140)`[^5]. 



    * [b. analisisfinal.py](https://github.com/blackcub3s/MSc-FinalThesis/blob/main/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/b.%20analisisfinal.py): The most important functions of this file are:

        `fes_dataframe_despres_dels_3_criteris_dinclusio(d)`: This function narrowed down the subjects that went into the machine learning models according to inclusion criteria of the methods section of my final thesis. 



        https://github.com/blackcub3s/MSc-FinalThesis/blob/3cfe02d11977db9d5d736267c5d6f6114fe82039/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/b.%20analisisfinal.py#L48-L243



        The most important inclusion criteria within that function was how I decided to filter out the subjects who would not convert to Alzheimer's (see the footnote to understand the reasoning that was to be followed[^6]). The idea was to choose the criterion of considering those who wouldn't turn to alzheimer's (MCI-nc) as those subjects who hadn't turn to alzheimers above the time period $t$, such that: $t > \mu + n\rho$

       

            

        Where  

        

            μ = 2.3 years (average time of turning to alzheimers for those who turn)

            ρ = standard deviation of those who turned to AD

            n = 1.85 (arbitrary number, but if we consider the distribution of time for MCIc to be normal, that would way above the central tendency of the data).



        In the following code, you can see I stored the subjects I would finally analyse within the `df_MCInc` dataframe, by writing this:



        https://github.com/blackcub3s/MSc-FinalThesis/blob/a2479224491c308d72a412e4dc2724107a830fe0/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/b.%20analisisfinal.py#L108-L111



        `mat_corr_labels_i_pipeline()`: This function took the numpy ndarray `arr_TS` with shape *(57,214,140)* (57 subjects with 214 ROIs each and 140 Time Series per ROI) applied the Functional Connectivity definition. The Functional Connectivity (*FC*) is a construct that tells how well connected a brain is. Here, the *FC* of one given subject was defined as all possible correlations between the time series of the subject's ROIs, vectorized in a single line. Namely, for each matrix  *(n,m) = (214,140)* we would get a *(n,n) = (214,214)* pearson correlation matrix (or the adjacency matrix of a complete graph with 214 nodes) called `arr_cor` and stored as [out_FC(57 subjectes).npy](https://github.com/blackcub3s/MSc-FinalThesis/blob/3cfe02d11977db9d5d736267c5d6f6114fe82039/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/out_FC%20(74%20subjectes).npy)[^7] as obtained in this lines of code:



        https://github.com/blackcub3s/MSc-FinalThesis/blob/3cfe02d11977db9d5d736267c5d6f6114fe82039/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/b.%20analisisfinal.py#L305-L321



        And after that, we would take the `arr_corr` for each $i$ subject in a for loop (`arr_corr[i]`) and obtain the flattened vector with the pearson correlations called `ll_vector` and stored within the `X` matrix that will go to the machine learning model input matrix. `ll_vector` will have shape $(1/2) \cdot n(n-1)$:



        https://github.com/blackcub3s/MSc-FinalThesis/blob/a2479224491c308d72a412e4dc2724107a830fe0/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/b.%20analisisfinal.py#L374-L379



        The auxiliary function call in line 378 that allows us to do obtain `ll_vector` (`ll_ROIxROI_flattened` within the function) is here: 



        https://github.com/blackcub3s/MSc-FinalThesis/blob/7592ad1c1d86a1427c29db471564e236368d259b/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/funcions_auxiliars.py#L29-L44



        Finally we obtain the *Y* matrix, with the labels (wheter or not the MCI will convert to AD or no, with a 1 or 0 respectively) and run the machine learning models:



        https://github.com/blackcub3s/MSc-FinalThesis/blob/60de1f5d9149d065bfdf3a592d007802fb63c3ec/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/b.%20analisisfinal.py#L403-L417



        `apc.PCA_10foldCV_ROC(X=X .. )` is the function call to the `PCA_10foldCV_ROC` of another file, which are described down below:



    * [aux__plot_roc_crossval.py](https://github.com/blackcub3s/MSc-FinalThesis/blob/7592ad1c1d86a1427c29db471564e236368d259b/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/aux__plot_roc_crossval.py): As we said, this is the file where the machine learning models are run. 10-fold-crossvalidation, PCA, Artificial Neural networks., etc. Within this file the most important function is the one we called at line 410 in the previous file above and goes as follows: 



        https://github.com/blackcub3s/MSc-FinalThesis/blob/7592ad1c1d86a1427c29db471564e236368d259b/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/aux__plot_roc_crossval.py#L64-L307



    * [aux__ploteja_matrius_correlacions.py](https://github.com/blackcub3s/MSc-FinalThesis/blob/main/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/aux__ploteja_matrius_correlacions.py)



    * [aux__reduccio_dimensions.py](https://github.com/blackcub3s/MSc-FinalThesis/blob/main/DataAnalysis/inContext/DADES_PREPROCESSADES_FINALS/aux__reduccio_dimensions.py): This file was done in order to implement a graphical interpretation for the explained variance of a principal components analysis (i.e it plots the eigenvalues of each component for the previously chosen number of components). The scree plot is a graphical way of seeing whether each component of a PCA adds value to what we are trying to model in terms of parsimony. Softwares such as SPSS or R have the possibility of getting a Scree plot easily, but python doesn't (at least, not while I was doing this thesis). So I coded this program to get the scree plot.

    



The packages I've used in this thesis are [sklearn](https://scikit-learn.org/stable/) for machine learning, [matplotlib](https://matplotlib.org/) and [seaborn](https://seaborn.pydata.org/) for graph generation and visualization, [pandas](https://pandas.pydata.org/) for data analysis, [numpy](https://numpy.org/) for treating the multidimensional arrays that came out of the NIFTI files that came out of the fMRIs.



[^1]: Lots of files are missing as github has limited space, so this is a simplification of the directories of my final thesis, not an exhaustive tour to it!



[^3]: NIFTI files are not here, as they occupy several Gigabytes of data. This is just for showcasing purposes.

[^4]: Region of Interest (ROI).

[^5]: `(93,214,140) --> (SUBJECTS, ROIs,TIME SERIES)`

[^6]: In my final thesis I took subjects that were at risk of developing Alzheimer's disease but didn't have the disease (these are what are called *Mild Cognitive Impairment* or MCI). At the moment of diagnosis as MCI, they had to undergo a brain fMRI scan. Then time went on and after some years some of them became Alzheimer's disease (MCI-c), and some didn't (MCI-nc). The ones who I labeled as MCI-c are clear to label, but the ones who are MCI-nc are not clear (because there might be not enough longitudinal data because the patient might die, get discontinued from the study). So it was important to choose a criterion to decide how many years of follow up are *enough* for a patient to be considered free of Alzheimer's. 

[^7]: The actual code got 74 subjects, as changes in the inclusion criteria (MCI-nc cutoff time point) varied.

[^8]: The [ADNI](https://adni.loni.usc.edu/)(Alzheimer's disease neuroimaging iniciative) gathered together high quality, free to use data on the alzheimer's disease, an without this data this thesis wouldn't have been possible.