Quest2Dx: Questionnaire Transformers for Parkinson’s Disease Screening

Introduction: Parkinson’s disease (PD) is a progressive neurological disorder that is typically not diagnosed until it has reached intermediate or late stages, primarily due to the subtlety of its early symptoms and the lack of early-stage recognition by primary care physicians. Consequently, patients often miss the opportunity for early intervention. The gold-standard methods for diagnosing using non-motor symptoms are invasive and, therefore, not ideal for early detection. Recently, data-driven approaches have shown promising results for PD diagnosis; however, they largely rely on primary care reports, which tend to focus on motor symptoms that manifest in later stages of PD.
Deep learning models, which could potentially better identify complex patterns indicative of early PD, struggle with the high volume of categorical data and missing entries typical of tabular datasets. Furthermore, missing entries in questionnaire data are a common problem that pose a key challenge for its analysis. The current strategies consist of simply removing the samples with missing entries or imputing them using statistical methods. This leads to degraded accuracy due to smaller datasets and noisy data. Our proposed method, Quest2Dx, aims to address the methodological and application challenges by proposing a novel approach for the early detection of PD from questionnaire data. The method leverages the powerful abilities of learnable embeddings in the transformer framework to represent the missing data entries and effectively model the key patterns in non-motor data collected from questionnaires.

Materials and Methods: Our proposed method analyzes tabular data collected from health questionnaires to differentiate responders with PD from healthy controls (HC). We employed the transformer network for this task, which takes representation tokens as input and outputs a class token. We encoded each entry of categorical response (e.g. 1, 2, 3, 4) into a token using an embedding layer. The embeddings map tokens to vectors of real numbers, capturing semantic information about the categories. In addition, we encoded the textual questions using an off-the-shelf large language model, BERT, to inform the meaning of the categorical response values. The output class token captures an overall representation of the subject and the patient classification is obtained by projecting that token with a fully connected layer. We further introduced a novel missing entry token to learn an embedding to represent the missing value. Figure 1 shows our overall framework.
We used data from the landmark dataset for PD, PPMI-online, which contains self-reported measurements collected through surveys of 40,012 subjects, including 27,802 PD and 12,210 HC. The data ranges from lifestyle to mental status and mood. We removed questions with >50% missing entries and features that report motor function decline , in order to reflect only early-stage indicators in our data. Answers for total of 131 questions were used.

Results, Conclusions, and Discussions: We evaluated the effectiveness of our proposed model at two levels. We first compared our model against other machine learning benchmarks, including Random Forest (RF), Support Vector Machine (SVM), XGBoost, Multi-layer Perceptron (MLP), and baseline Transformer. To validate the effect of the proposed missing data handling strategy, we compared the performance of transformers under no imputation (sample removal), mode-based imputation, and missing token representation.
Our proposed method, implementing the missing entry special token and question encoding with BERT, was able to consistently achieve the best performance as seen in Figure 2. The deep learning frameworks showed better performance than the traditional machine learning methods, with the transformer model performing the best. The proposed special token for missing entries contributed to the high performance compared to simple data imputation, at both lower (20%) and higher (50%) schedules of missing entries. The data imputation strategy balances a tradeoff between noise and dataset size. The lower schedule of missing entries will have less noise introduced from filling out the blanks in the dataset; however, it will have less samples due to strict 20% missing entries maximum constraint. In this dataset, no steep accuracy degradation was seen for any model, and Quest2Dx still remains the best performing model at 50% schedule. Showing its robust performance regardless of the number of missing entries.
Our proposed framework was able to effectively handle missing entries in increasing percentages. That will greatly increase the applicability of the method, as many of the self-reported questionnaire datasets suffer the problem. Our method will enable researchers to utilize such incomplete datasets in their research. More importantly, our proposed method was able to effectively classify PD patients using only self-reported data, which shows the promise of screening large population at primary care level for early detection of PD. That in turn will help improve the healthcare outcomes and lower the cost.