【无标题】

Hello everyone, my name is Du Xingze and today I would like to share with you a disease prediction model, it is a fusion model and composed of three parts, CNN-BiLSTM-Attention, .

222

In this talk, we will cover three main aspects: Firstly, it is the introduction of the background. Secondly, we will delve into how this model is constructed and finally, we will evaluate this model using various performance metrics.

333

By the year 2022, 537 million adults around the world were afflicted with diabetes, and this number is projected to climb to 643 million by the year 2030, marking a significant upward trajectory.

Additionally, the chart on the right underscores "China has the highest number of diabetes patients in the world.".，

Thus, early prevention becomes not just a suggestion but a critical necessity.

There are many methods to do this ,

However, traditional machine learning methods have certain limitations and struggle to effectively process the vast amount of data.  standalone neural networks also have their inherent shortcomings. It is simple for beginners to try fusion methods. This fusion aims to address the constraints present in both traditional machine learning and standalone neural network methods.

444

Now, let's talk about the first network we're looking to bring into the mix – the Convolutional Neural Network, or CNN. Why do I want to start with this one? Well, CNNs have this amazing ability to crunch data by using their special convolutional modules. And here's the cool part: by adjusting the size of its "filters" and how it moves through the data, we can fine-tune its performance and make it work even better for our needs.

555

Next up, we've got the Bidirectional Long Short-Term Memory, or BILSTM for short. Let's take a quick look at the basics of an LSTM unit. It's like a little memory cell that keeps updating itself. You can picture it as this chain of connections and formulas on the right that show how it all happens. It's all about understanding patterns and relationships in the data – a crucial piece of our puzzle.

666

Multiple LSTM cells can be chained together to form an LSTM layer that captures long-term dependencies in data.

But what about future information? That's where BiLSTM comes in. By stitching together two LSTM chains, we can utilize both past and future information to improve our predictions.

777

In the end of the model ,it is the attention mechanism, This technique is based on human visual attention, where our eyes tend to focus on areas of interest. In a neural network model, simply, this means assigning higher weights to more important parts of the data when calculating results.

888

Now, we have learned the three parts of the cba model ,CNN ,Bilstm, Attention, we combine these three parts in a specific order. First, the CNN extracts local features, after that, the BiLSTM extracts sequence features that depend on both past and future data. Finally, the attention mechanism dynamically assigns weights to the results.

And that is a clear picture .

999

Now, let's move on to applying our model in real-world scenarios. This involves a few practical steps that tie everything together. We start with data preprocessing – getting our data all cleaned up and ready for action.

With that understanding, we roll up our sleeves and start coding. Using the structure we've just discussed, we build the model. It's like putting the pieces of a puzzle together – each module playing its role in this larger scheme.

During training, we're tweaking things based on the loss we're seeing. Our goal here is to find the best parameters that make the model work like a charm.

000

And then, the moment of truth – performance evaluation. We compare our model to others, and guess what? Our model shines brighter. We've even thrown in the MCC metric to make it more comprehensive. Providing a holistic assessment of the overall performance of the binary classification model. Compared to traditional models, our accuracy soars higher. And when we put it side by side with neural network models, ours comes out on top.

1111

The fusion model can leverage the advantages of both CNN  and BiLSTM to utilize the short-term and long-term features of the data.