Maddie Lykouretzos '23
Abstract
Schizophrenia is a complex disorder affecting 20 million people globally, with a drug market worth 6.8 billion USD in 2016. With the disorder manifesting itself with a variety of symptoms that differ from patient to patient, it is difficult for doctors to properly identify what treatments will work best, and for companies to find Schizophrenic patients whose subtype is applicable for their preclinical drug testing. The result is a failure to adequately treat a patient, and significantly less accurate results in preclinical trials. With a disorder that spans such a large range of causes and symptoms, it is no wonder that so many mistakes are made with diagnosing. In turn, companies have begun to use AI as a means of handling large amounts of data such as patient blood samples, lifestyle, and environment as input to predict the functional outcomes of patients with Schizophrenia. From these accurate predictions, doctors will spend less time, money, and effort into trial and error to find which drugs or therapies will be most beneficial to their patients. Additionally, companies will have higher success rates when testing drugs on specific Schizophrenia subtypes since predictive models will be able to accurately sort patients and predict the course of the disease. Overall, using AI to predict the functional outcomes of Schizophrenia is a massive leap towards revolutionizing the way these patients are treated, diagnosed and categorized.
Introduction
Precision psychiatry is a relatively new study at the junction of precision medicine, psychiatry and technology, creating a platform for doctors to target genetic biomarkers that are directly involved in altering a patient’s condition. Through this method, patients are able to receive accurate medical feedback on what available therapies or drugs will work best based on what the algorithm says. Often the method of precision psychiatry involves taking into account factors beyond what genetic biomarkers a patient exhibits, and expands to their environment and lifestyle. This method is being considered more and more for patients with mood disorders due to its ability to synthesize all of the different inputs and produce accurate predictions of how their already diagnosed condition will behave or how likely a patient is to get the disease.
Scientists Eugene Lin, Chieh-Hsin Lin, and Hsien-Yuan Lane discovered that by using a bagging ensemble machine learning method with an M5 Prime feature selection, they were able to accurately predict the functional outcomes of patients with Schizophrenia. Schizophrenia is often differentiated among patients through the use of the Quality of Life Scale (QLS), a tool that allows doctors to assess a Schizophrenic patient’s functional outcome through factors such as curiosity, empathy, sense of purpose, social activity, social withdrawal and anhedonia. Doctors can also assess a patient’s Schizophrenia by using the Global Assessment of Functioning (GAF) to evaluate social, occupational, and psychological functioning.
Methods
The program used by the scientists includes an M5 Prime algorithm (M5P) which utilizes conventional decision trees and linear regression to detect the presence of single nucleotide polymorphisms (SNPs) in the patient’s blood. SNPs can sometimes be harmless but are often linked to disease in humans. The eleven SNPs found, have been chosen by the scientists due to their correspondence to Schizophrenia from previous studies and their presence in the 302 patients used in the study highlighted in this particular article. The genotype frequencies of each of these biomarkers are determined by the M5P algorithm and are then assessed for Hardy-Weinberg equilibrium to ensure that they are present in the patient’s blood. Once the biomarkers in the patient are identified, they serve as input for the bagging ensemble machine learning method, where bootstrap reproduction is used to generate a number of versions of a predictive model in order to find the model with the best performance. The result is an accurate prediction of the patient’s functional outcome, in other words, an assessment of how their condition behaves and affects their daily life based on the QLS and GAF. From these results, doctors can find the correct treatment and or therapy that will best suit their patient.
A study using this predictive method was conducted on 302 Schizophrenia patients aged 18-65 years from the China Medical University Hospital and affiliated Taichung Chin-Ho Hospital in Taiwan. The patients were otherwise physically healthy. Venous blood was taken from each patient to check for the biomarkers and their lifestyle and environment were also taken into account. The functional outcomes were measured based on each patient’s responses to the QLS and the GAF. Different state of the art tests such as random forests, Linear Regression, Support vector machine, and Multilayer feedforward neural networks were used as independent variables to see how accurately they compared to the M5P with bagging ensemble, all based on the patient’s QLS and GAF responses.
Results
The results showed that the bagging ensemble alone was more accurate than other methods in predicting a patient’s QLS, and was even more accurate when paired with the M5 Prime feature selection.While the bagging ensemble alone was not the best in predicting GAF, it became significantly more accurate than other methods when using the M5P selection. In short, there is sufficient evidence to show that this method is not only successful, but outperforms and gives better results than any other known methods. Not only that, but the bagging ensemble with M5P needed less input than the other methods and still was more accurate.
Discussion
There are several other AI methods that are used in precision psychiatry: Logistic regression, Naive Bayes, C4.5 decision tree, Linear regression, Support vector machine, Multilayer feedforward neural networks, and Random forests. All of these methods take an input to predict the functional outcome of a patient and therefore help doctors decide what kind of treatment or medicine will be most beneficial. While the bagging ensemble with feature selection does not differ in this goal, it does differ in its method and result. This method takes advantage of using biomarkers, and does so much more efficiently than other methods since it utilizes the M5P feature selection. When paired with the M5P feature selection, the bagging ensemble program becomes significantly better at predicting the Quality of Life Scale (QLS) and the Global Assessment of Function (GAF) in patients than other predictive models.
Due to the accuracy and high rate of success, this method has the potential to not only be the leading predictive AI model for Schizophrenia patients, but to expand to an array of other mood disorders and mental illnesses such as Major Depressive Disorder and Bipolar Disorder. By implementing this technology in clinics, it can be used to accurately examine how a patient’s disorder will affect them and their daily life based on the Quality of Life Scale and Global Assessment of Functioning. The AI will also be incredibly useful in recruiting patients to see if their subtype of a disease is applicable for preclinical trials and drug testing, which will overall save the company money and lead to significantly more accurate results in their trials.
Since this technological discovery is relatively new, the authors of the article have a number of steps to take before their AI program can be used by clinics and companies. Getting a patent on the predictive method and its code is essential before possibly building a company around it. The next step is to contact other biotech companies and use the AI program to sort patients involved in preclinical testing for Schizophrenia drugs. Some of the major players in the Schizophrenia drug market are Johnson & Johnson, Bristol-Myers Squibb and Alkermes. Another notable player is Eli Lilly, which has a dedicated research budget to developing drugs for the complex disorder. However, phases II and III of their drug, pomaglumetad methionil, ultimately failed. The problem in these trials wasn’t the drug itself, as suggested by preliminary research, but rather the patient population used in the trials. In the future, bagging ensemble machine learning with M5P feature selection can prevent this from happening by partnering with these major companies to correctly sort patients that are applicable for drug trials. From here, the AI program can expand to test other diseases and search for biomarkers that correspond to the patients involved.
To conclude, using a bagging ensemble machine learning method with an M5 Prime feature selection predicts the functional outcome of Schizophrenia patients better than any other known method and therefore can be applied to a variety of other mood disorders. This saves doctors, patients, and companies a massive amount of time, effort, and money.
Bibliography
Breiman, L., Bagging predictors, Machine learning, 24, 123-140, (1996) · Zbl 0858.68080
Emamian ES, Hall D, Birnbaum MJ, Karayiorgou M, Gogos JA. Convergent evidence for impaired AKT1-GSK3beta signaling in schizophrenia. Nat Genet. 2004 Feb;36(2):131-7. doi: 10.1038/ng1296. Epub 2004 Jan 25. PMID: 14745448.
Fernandes, B.S., Williams, L.M., Steiner, J. et al. The new field of ‘precision psychiatry’. BMC Med 15, 80 (2017). https://doi.org/10.1186/s12916-017-0849-x
Lin CH, Huang CL, Chang YC, Chen PW, Lin CY, Tsai GE, Lane HY. Clinical symptoms, mainly negative symptoms, mediate the influence of neurocognition and social cognition on functional outcome of schizophrenia. Schizophr Res. 2013 May;146(1-3):231-7. doi: 10.1016/j.schres.2013.02.009. Epub 2013 Mar 9. PMID: 23478155.
Lin E, Lin CH, Hung CC, Lane HY. An Ensemble Approach to Predict Schizophrenia Using Protein Data in the N-methyl-D-Aspartate Receptor (NMDAR) and Tryptophan Catabolic Pathways. Front Bioeng Biotechnol. 2020 Jun 4;8:569. doi: 10.3389/fbioe.2020.00569. PMID: 32582679; PMCID: PMC7287032.
Lin, E., Lin, CH. & Lane, HY. Prediction of functional outcomes of schizophrenia with genetic biomarkers using a bagging ensemble machine learning method with feature selection. Sci Rep 11, 10179 (2021). https://doi.org/10.1038/s41598-021-89540-6