Implementation of classic machine learning with Dräger

The world of large language models – LLMs – is evolving steadily, with new models and tools being released almost daily to improve the quality of results and open up new use cases. These models, which have been developed with gigantic investments in computing power and training data preparation, have made artificial intelligence a tool for everyone. They are particularly well suited for natural language processing, sentiment analysis and language translation. Currently, the main use cases are code and text generation and summarization, image analysis and image and video generation. In conjunction with agents, i.e. systems that can perform tasks and make decisions independently, they have the potential to automate and transform business processes.

LLMs are designed to process unstructured data. They can recognize patterns and relationships in the data without being explicitly told what to look for. These models form a special part of the broad spectrum of machine learning, that has been pushed very far and will be pushed further. In contrast, traditional machine learning algorithms require labeled training data and often rely on human experts to define the features to be used for analysis. The classic use cases, for example for the purpose of recognition and prediction, remain important. In our most recent project with our customer Dräger, we used classic machine learning algorithms to make predictions based on medical data.

What is classic machine learning?

AI pioneer Arthur Samuel provided a very fitting definition of the term back in the 1950s: “the field of study that gives computers the ability to learn without explicitly being programmed.”

In some cases where something needs to be automated, programming is too time-consuming or not possible at all. For example, when it comes to recognizing something in images. While humans can easily manage this task, it is difficult to develop an algorithm for it. Instead, the machine learning approach is that the computer learns through experience. Achieving reliable results along this path requires a large base of training data, which is used to train a so-called neural network. This is a network of artificial neurons that is modeled on a brain. The neurons are parameterized during training so that their weights and bias values are optimized by repeated adjustments to the training data in order to deliver the desired results for all inputs as far as possible. These results must be defined beforehand and assigned to the data. The aim is for the network trained in this way to deliver the correct results even for input data that is unknown to it.

The challenge: clarity in alarm management

Intensive care units are highly complex working environments in which patients are continuously monitored. This monitoring generates numerous alarms of varying priority. Life-threatening alarms can get lost in the multitude of less critical alerts – a challenge for both nursing staff and patient safety.

Together with our customer Dräger, we asked ourselves whether it was possible to predict whether a life-threatening alarm would soon be triggered at a bed based on the alarms that had already been triggered. The aim was to develop a solution that works preventively and can still be integrated into existing systems. Real-time capability was a key objective to ensure that the forecasts were available immediately.

The data: the basis for predictions

The project was based on an extensive alarm database. This data was collected and provided by Dräger’s devices in intensive care units. Each data set included parameters such as alarm level, trigger time, alarm duration, measured value and the corresponding threshold value. This data had to be carefully prepared in order to serve as a training basis for the model. 

In order to make the data usable for the machine learning model, several data preparation steps were necessary:

1. Filtering irrelevant information
   • Alarms that were triggered by technical errors or a missing connection to the patient were removed as they have no added medical value.
   • Multiple entries of the same alarm were removed to keep the database consistent.
   • The cleansing and preparation processes resulted in a reduction of the database of about 50%.
2. Normalization and transformation
   •  Numerical values such as measurement parameters and alarm duration were scaled to a value range between 0 and 1 using min-max normalization to facilitate model training.
   • Categorical variables such as alarm levels or parameter units were converted into numerical representations using one-hot encoding.
3. Sequencing
   • Since alarm prediction is a temporal problem, alarms were grouped into sequential groups. These sequences represent the alarms of a single bed over a defined period of time.
4. Target definition
   • The aim of the model was to predict whether a life-threatening alarm will occur within a sequence.
     • For this purpose, the alarm sequences were randomly shortened.
     • Alarm A, which would have followed immediately after the shortened sequence, served as the target for the prediction.
     • If alarm A was a life-threatening alarm, the model should predict a life-threatening alarm based on the remaining sequence.
     • On the other hand, if A was not a life-threatening alarm, the model should not predict a life-threatening alarm.

Example sequence without normalization before target definition

Alarm level	Parameter-ID	Parameter value	Parameter unit	Parameter upper limit	Parameter lower limit	Alarm duration
SER	ART D	220	MmHg	110	50	20
SER	SpO2	85	%	100	90	22
SER	ART D	160	MmHg	110	50	25
L-T	VT	VT	0	0	0	19
L-T	VF	VF	0	0	0	10
ADV	ART S	163	MmHg	160	90	30

Example sequence without normalization according to target definition. Target Prediction of a life-threatening alarm characterized by alarm level L-T (life-threatening)

Alarm level	Parameter-ID	Parameter value	Parameter unit	Parameter upper limit	Parameter lower limit	Alarm duration
SER	ART D	220	MmHg	110	50	20
SER	SpO2	85	%	100	90	22
SER	ART D	160	MmHg	110	50	25

Challenges with the database

Despite the extensive database, the data set also had weaknesses:

• No patient-specific information such as age, gender, previous illnesses or vital signs such as pulse were available. These could provide important additional characteristics for more precise predictions.
• The data structure was based solely on alarm events, meaning that the model had no direct insight into the clinical context or the patient’s condition.

The quality and structure of the data were critical to the model’s performance. The processed data provided a solid foundation for building a proof-of-concept model, but also showed the potential for future improvements through expanded data sources.

At the end of this process, a significantly reduced, structured data set was suitable for training a model.

The solution: a model for specific predictions

Artificial neural networks (ANNs) were deliberately used for the project, as they are particularly well suited to recognizing complex patterns and relationships in the data. The flexibility of ANNs made it possible to achieve precise results even with less structured or high-dimensional data sets.

Approach:

• Feature extraction: Input features were extracted from the sequenced alarms, including alarm levels, measured values and alarm duration.
• Classification model: The model was trained to predict whether a life-threatening alarm would occur based on a sequence.

A particular challenge was to train the model to make meaningful predictions without directly accessing the life-threatening alarms that occur later in the sequence. For this purpose, the corresponding alarms were removed and only the information that the sequence led to life-threatening alarms was preserved. 

Results and practical application

The model showed a prediction accuracy of up to 85% in the validation phase. This accuracy does not allow for practical application, as hospital staff must be able to trust the prediction completely in order to be able to make action decisions based on it. However, there is potential for improvement, for example by incorporating real patient data such as age, gender, previous illnesses or even pulse into the neural network as input and not just alarms. 

The solution was integrated as a proof of concept so that it could be integrated into existing systems via an interface in Microsoft Azure. Azure helps to ensure the desired real-time capability.

However, the project also showed that classic machine learning methods are only part of the solution. The quality of the results depends largely on the preparation and structuring of the data. The task of involving nursing staff in the interpretation and utilization of the predictions also remains an important challenge.

Conclusion: Realistic applications of classic approaches

The project  with Dräger makes it clear that classic machine learning algorithms can be used in a targeted manner in specific applications. However, it also showed that such approaches have their limits, especially if the available database is only of limited significance. The focus was on a pragmatic solution that utilized the available data and was intended to supplement existing workflows, although the prediction accuracy of up to 85% achieved is not yet sufficient for a practical application.

This makes it clear that traditional AI approaches can continue to make valuable contributions to modern medical technology – especially as a building block in development processes or proof-of-concept projects. However, extended data sources and advanced models are required for use in everyday clinical practice in order to meet the high demands for precision and reliability.

With projects like this, we at akquinet tech@spree GmbH emphasize our ability to develop innovative and practical solutions. Classic machine learning is more than just an alternative – it is an integral part of our modern technology toolbox.