Analysis of fmri data of depressed patients using ml

Inspiration

Clinical depression is an extremely common mental health disorder which is characterized by decreased enjoyment of daily activities and a loss of purpose in life. It is a serious condition which affects an estimated 19 million American adults. Depression left untreated can impair a person’s ability to function in daily life and in extreme cases may lead to suicidal behaviors. It is thus essential that medical professionals are able to identify and successfully treat depressed patients.

The most effective treatments for depression are cognitive behavioral therapy and medication. CBT is a form of treatment in which patients work to a professional therapist to change their perceptions. The therapist identifies a patient’s harmful thoughts, assesses whether they are accurate, and if not encourages the patient to modify them using specific strategies. An alternative treatment is the use of prescription drugs, the most common of which are selective serotonin reuptake inhibitors. SSRIs increase the brain’s uptake of Serotonin, a neurotransmitter found in the central nervous system that regulates mood and behavior.

Neither CBT nor any prescription medication works perfectly for all patients. Depending on the individual, certain treatments may work better than others. In the status quo, psychiatrists struggle to determine the right treatment for each patient, and thus treat patients with trial and error. It can take several years for a psychiatrist to find adequate treatment for a patient. The long gap between onset of depression and its treatment is unacceptable. If depression is left untreated, the symptoms can worsen and cause severe side effects: it can ruin relationships, lead to drug addiction, and even infliction of self-harm. In order to avoid these effects, psychiatrists must be able to immediately identify the best treatment for each individual. Depression treatment can be made more effective if a machine learning algorithm can predict an individuals response to the treatment before it is prescribed to them.

What it does

It predicts whether a depressed patient will respond positively to a certain prescription medication with a confidence of above 65%.

How we built it

The clinical study which this paper obtains its data from only imaged a couple hundred patients, making it difficult to properly train a machine learning algorithm to predict a depressed individuals response to a prescription drug. In an attempt to address this challenge, the data was split into training and test data and inputted into the algorithm in this order. The algorithm uses the training data to make better predictions and the test data to check whether its predictions are applicable to similar data. By evaluating the algorithm’s accuracy in classifying the test data, it can be determined to what extent the algorithm will work in a broader context. Several statistical measures were computed to measure its accuracy, most notably the receiver operating characteristic curve. The ratio of the true positives to the false positives was compared with that of the average psychiatrist’s successful 1st prescriptions to unsuccessful 1st prescriptions to determine the benefits of this approach.

Challenges we ran into

Two main challenges must be overcome before a successful machine learning algorithm can be implemented. First, the 4 dimensional fmri data must be processed into 2 dimensional correlation matrices before it can be inputted into a machine learning algorithm. These matrices measure the similarities in brain signals over time between two ROI’s and thus evince a connection between them. Second, the algorithm must avoid overfitting its data if it is to be useful in practical applications.

The first challenge is relatively easy to tackle due to the abundance of open-source image processing technology, most notably nilearn, a Python module which enables the user to apply advanced machine learning and multivariate statistical techniques on neuroimaging data. The data and msdl atlas were loaded, and patients’ time series data were extracted by instantiating a nilearn.input_data.NiftiMapsMasker object and passing it the file name and atlas of each file. Patients’ region signals were extracted by calling the fit_transform method on the masker object and passing it the file name. Simultaneously, each patients’ response to the treatment is retrieved and stored as a label. Connectivity matrices were calculated by importing ConnectivityMeasure from nilearn.connectome, specifying the type of correlation, and calling the fit_transform method while passing it the patients’ region signals. The resulting matrices will be key to determining whether a patient will show a significant response to treatment.

Accomplishments that we're proud of

We achieved a classification accuracy of 67% for placebo patients by inputting tangent_correlation_matrices into our machine learning algorithm. We achieved a classification accuracy of 65% for SSRI patients through the same input.

What we learned

Machine learning can serve as a supplement to a psychiatrist's opinion as to which prescription drug to treat a patient with.

What's next for family_depression

We want to get more functional magnetic resonance imaging data from depressed patients so that we can better train our algorithm to classify a patient's response. We also want to implement ensemble learning to our model because it leverages the power of multiple weak algorithms to create a stronger one.