Still can't understand the doctor's handwritten prescription? Alibaba's new technology can help

Currently, many medical records have been converted into electronic medical records (EMR), and information in traditional paper medical records can also be extracted through image character recognition technologies such as optical character recognition (OCR).

Traditional medical record analysis is based on the doctor's experience to manually understand and analyze the information in the medical records. In some simple cases, preliminary analysis of medical records can be performed using artificial intelligence (AI) technology, automated intelligent operations, or input provided by doctors. This is also the medical automation technology being implemented in some large hospitals.

Among these, the recognition and classification of medical records are also regarded as an important part of realizing the automation of medical processes. However, some medical records may include handwriting errors, typing errors or some newly emerging terms, which are difficult to recognize and process for traditional OCR processing technology. It is precisely because of this that there has been a metaphor among the people that doctors' handwritten medical records are like "ghost paintings".

To this end, Alibaba applied for an invention patent called "Identification of Entities in Electronic Medical Records" (application number: 201980003795.2) on April 25, 2019, and the applicant was Alibaba Group Holding Limited.

Based on the information currently disclosed in the patent, let us take a look at this automatic recognition technology for medical records.

As shown above, it is a block diagram of an example of a training process 100 for identifying and classifying terms invented in the patent. This framework can be used to create entity recognition training and identify electronic medical records. First, training text is extracted from training data, and the training data includes electronic medical records. The training text is the medical diagnosis text in the electronic medical record.

Secondly, the texts need to be divided. The purpose of dividing the texts is to classify some medical terms. In this way, when the algorithm automatically classifies the texts, it can use the characteristics of different word classes, so that the machine can simulate human analysis of medical records. Of course, this process requires word vector training. In the field of OCR, commonly used training methods include the cw2vec algorithm and the BiLSTM-CFR training model.

If you are familiar with artificial intelligence algorithms, you must be familiar with the LSTM algorithm, which is a special RNN network that is mainly used to solve the gradient vanishing and gradient exploding problems in the long sequence training process. It also has a good effect on text interpretation work such as sentence understanding. Therefore, with the help of these algorithms, word annotation information and training word vectors can be learned, and the learned models can be stored for automatic recognition of medical records next time.

The figure above shows a predictive text extraction process 200 for identifying and classifying terms, which is carried out after performing the above-mentioned identification and classification term training, and can also improve the process 100. This predictive process first needs to extract text, which also comes from electronic medical records or physical medical records.

Different from process 100, process 200 adds the learning of new words and vocabularies, that is, for words that have appeared in the training data, the word vectors can be used directly, but if new words and vocabularies that have never appeared before appear, the new words can be calculated from the new word vocabulary 208 and the stroke vector 210, and the new words are decomposed into stroke sequence and the strokes of the new words and the new word vectors are obtained through a sliding window.

With such a mechanism in place, the system can ensure that it can not only recognize medical records that have already been recorded, but can also automatically perform text recognition for new medical records, thereby automatically converting doctors' obscure fonts into easy-to-understand electronic information.

The above is the medical record recognition invented by Alibaba. This method uses the artificial intelligence algorithm BiLSTM-CFR to train the model, so that it can recognize the text in the medical record. It has a good recognition effect for both handwritten and electronic medical records. Such technology can not only improve the automation level of the hospital, but also avoid errors caused by the difficulty of recognizing handwritten medical records.