Data quality is a determining factor for the performance of artificial intelligence algorithms. If data contain errors, have missing values, or use inconsistent formats, the model may generate inaccurate results or even stop working properly. Quality is not ensured only during analysis: it starts at the point of data collection.

In contexts such as digital health, this issue is especially relevant. An incorrect measurement, an empty field, or a wrongly recorded unit can affect patient analysis and reduce the reliability of predictions.

Principles for ensuring data quality

To ensure high data quality, it is necessary to adhere to a number of principles:

• Accuracy: record correct values and minimise errors. For example, if a monitor shows a heart rate of 120 beats per minute, the recorded value must be 120.
• Avoiding missing data: prevent empty fields and absent records. In the case of a sepsis patient, variables such as heart rate, blood pressure, temperature, or weight may be essential for the model.
• Consistency: use the same units and formats across all sources. Recording weight as “70 kg” at one time and “154 lb” at another can cause confusion and errors in the analysis.
• Relevance: collect only the data necessary for the analysis objective. Information such as a patient’s eye colour adds no value to a sepsis risk model.
• Validity: check that values comply with defined rules and ranges. A body temperature of 50 °C should be flagged as an invalid value.
• Uniqueness: remove duplicate records to avoid giving a single case artificial weight in the analysis.
• Precision: maintain an appropriate level of detail in numerical data. Recording weight as “68.5 kg” may be more appropriate than always rounding to “68 kg”.
• Consolidation: uniformly integrate data from different sources, such as sensors measuring different variables from the same patient.

The impact of poor-quality data

Incomplete data can lead to the loss of important variables for the model. If a column contains too many missing values, it may become useless for analysis.

It is also necessary to identify constant data. When a variable always shows the same value, the algorithm cannot extract useful information from it. A variable with no variation does not help distinguish between different cases.

When these problems accumulate, no algorithm can fully compensate for the lack of quality in the input data.

Applying these criteria from the point of collection enables the development of more accurate models and better-informed decisions. In short, data quality directly conditions the quality of the models and the decisions based on them.