A large part of medical practice involves pattern recognition. A clinician may note that a few key aspects of a patient’s history, their demographic data, their clinical examination, and chest radiograph fit a pattern they recognise as making a particular diagnosis highly probable. This pattern involves more than one domain of data acquisition, and both within and between these domains our ability to recognise patterns may be affected by how the information is presented to us. If data are presented to us verbally, the ordering of this information may be crucial. For example, verbal instructions on how to get from ward A to ward B are easier to understand and use if they are given in consecutive order starting from ward A and ending up at ward B rather than the instructions coming in random order. The order in which a patient’s history is presented to a colleague is an example of this. The graphical presentation of tables of numbers may improve the usefulness of the data,¹ especially if there is a shape in the data that conveys the signal and the time required to search the data for any signal is thereby reduced.² In this issue of Thorax Reddel and colleagues³ question whether we are doing enough to present our patients’ peak expiratory flow (PEF) data in a manner that is likely to facilitate both clinicians and patients distinguishing the signal or message in the data from all the noise.

Research continues to add to our understanding of how the brain detects and learns patterns. Facial recognition and the interpretation of facial expression are key aspects of human interaction involving complex pattern recognition. The ability of a clinician to interpret subtle changes in facial expression of patients during a consultation is important to ensure optimum communication has been achieved. Functional imaging has shown that specific areas of the brain—the fusiform face area on the inferior surface of the temporal lobe and the occipital face area—are involved in face recognition.⁴ Recent functional magnetic resonance evidence indicates that these areas appear to be used for expert or higher level recognition irrespective of the type of image under consideration.⁵ Research into whole person recognition, which involves the assessment of the face and body habitus, has found that the ability to recognise people was not impaired by showing the subjects in different postures whereas changes in the subjects’ clothing did impair it.⁶ This indicates that related visual effects can be distracting and alter our ability to recognise patterns in complex visual data. So, when constructing charts or graphs to display data, we must be careful that all non-essential items (“chart junk”) are left out. The trend to use three dimenstional graphics where two dimensional would suffice is an example of unnecessary distracting information; the exact projection of the top of a three dimensional bar or column onto the relevant axis is often distorted by the three dimensional effect. A recent survey of American medical journals found that, in 74 pharmaceutical graphical presentations, 66% contained “chart junk”, 46% had redundancy in the presentation, and over a third had numerical distortion.⁷

PEF variability is a helpful signal in the diagnosis and management of asthma. If PEF readings are presented as just a stream of numbers on a sheet of paper, which some patients will offer up in the clinic, it is a laborious process to check through these to find out what is going on. Graphical presentation of the data will help, but it has been shown that subjects serially process only fixed amounts of data at a time and so some types of graphical display are better than others in getting the information across quickly.⁸ The way data are presented on graphs or diagrams can therefore influence our ability to spot any signal. Reddel and colleagues³ point out that the abscissa scale for PEF plots has not been standardised. They found that each of 17 different PEF charts they obtained had slightly different scales. They have indicated that, if PEF data are plotted on a compressed time scale, then the ability to detect a true change in PEF is enhanced. Another example of this effect of scaling in respiratory practice is the presentation of flow-volume loops where incorrect scales will distort the data and may falsely suggest to the observer the presence of upper airway obstruction. The ATS document for standardising spirometry indicates the optimum scaling for presenting flow-volume loops (2 l/s of flow against 1 litre of volume) to avoid this sort of error.⁹

Computers can be used to help identify patterns in data by the training of neural networks and this approach has been successfully applied to the recognition of upper airway obstruction from flow-volume curves.¹⁰ For PEF data, discriminant analysis was employed by Gannon et al¹¹ to facilitate the recognition of work related changes in PEF and their computer program has led to improved sensitivity and specificity when diagnosing occupational asthma. Statistical process control (SPC) techniques¹² have been applied to PEF data as a means for improving the detection of true signals in the data. This technique assumes that the data should vary randomly following a Gaussian distribution and so newly acquired data points are compared with an estimate of the usual level of variation about the mean, which is the standard deviation (SD) of the data. A significant deviation from the baseline state is judged to have occurred if a single data point is more than 3 SD from the mean, or two of three sequential data are between 2 and 3 SD from the mean, or four of five sequential data are between 1 and 2 SD, or eight successive data are between 0 and 1 SD from the mean.¹² The assumption of a Gaussian distribution may be true for once a day morning pre-bronchodilator PEF values and this technique has been successfully applied for the detection of changes from baseline in otherwise stable patients.¹³ If several data points a day are used for asthmatics who have a large morning dip in PEF, then this assumption may no longer be correct as there are, in fact, two populations of results being lumped together. The application of SPC may not therefore be so helpful for other applications of PEF data, such as the diagnosis or monitoring of recovery from an acute attack.

The above control charting method requires an estimate of the baseline mean value and its variance which are hard to obtain from hand recorded data. The use of data logging meters facilitates this approach and they are now available for under £10 (15 euros), with associated PC software about twice this cost. The micro chips in these devices usually have spare capacity, so interactive meters that are tailored to the subject and their requirements are entirely feasible. The use of SPC on daily PEF values could easily be implemented and, if threshold criteria as above are breeched, then patients could be prompted to change their treatment and timing of PEF measurements to ensure control of their asthma is rapidly regained. The optimum abscissa scale for PEF charting will therefore need to vary according to purpose. For assessing stable asthmatics with a control chart, a once a day reading is needed and a compressed time scale is ideal. Reddel and colleagues have previously proposed an alternative method using the lowest morning pre-bronchodilator PEF in the week expressed as a percentage of recent best or predicted as the preferred indicator of control.¹⁴ However, in occupational asthma readings of up to every 2 hours may be needed to detect a work related change in lung function,¹⁵ and this frequency may initially be the minimum required to monitor recovery from a severe asthma attack.

The choice of abscissa scale may not be the only issue. An abstract some years ago at a British Thoracic Society meeting suggested that a log ordinate scale for PEF might be an improvement. The degree of variation in PEF is the main point of interest for helping with the diagnosis of asthma, and a 20% change from best value has been found to be associated with asthma.^16,17 Using a log scale for PEF might facilitate recognising this from the data. For example, in fig 1 there are three sets of identically patterned data each with different mean PEF. Plotting these data with PEF on a log scale as in fig 2 makes it more obvious that these all have identical percentage variation.

• Download figure
• Open in new tab
• Download powerpoint

Figure 1

 PEF data for three patients on a linear ordinate scale.

• Download figure
• Open in new tab
• Download powerpoint

Figure 2

 The same PEF data as in fig 1 on a log ordinate scale.

We must therefore be sure that we are always presenting PEF data (or any other data) to best advantage. As Reddel and colleagues argue,³ if we do not do this for our asthma patients then we let them down and undermine their efforts to help in the monitoring of their condition. In future, cheap electronic peak flow meters with data logging capability will increasingly be used and the graphical presentation of the data can be flexible and optimised for the purpose, be it once a day readings for stable monitoring, 4 times a day monitoring of recovery from an acute exacerbation, or 2 hourly or more frequent readings for occupational settings and severe attacks. The respiratory community should now work to determine the best scaling for these various purposes together with the best algorithms for detecting true changes in the PEF data. Only then will patients and clinicians get the most benefit from these data.