Facebook YouTube Twitter Linkedin
Facebook YouTube Twitter Linkedin


4 November 2020

Automated administration of Serial Subtraction in a remote data collection context: novel timing features related to task difficulty and participant demographics

New research presented at CTAD 2020 reports the successful derivation of novel measures of timing for a more nuanced assessment of performance on the Serial Subtraction task. 


  • Serial Subtraction is widely considered to be a valid and sensitive measure of processing speed and attention.
  • Variants of this task have been used in as part of neuropsychological screening tools (e.g. MoCA), and in conjunction with measures of gait and balance in the context of measurement of dual tasking, where Serial Subtraction is considered a means of applying cognitive load.
  • Despite the wide-spread utility of this task, measures derived from it have typically been limited to accuracy, or number of attempts over a given period (rate), and delivery has been restricted to in-person testing.
  • Here, we report on the derivation of novel measures related to the timing of responses from Serial Subtraction data collected in an automated and remote context.


  • To validate the automatic administration of serial subtraction using the Neurovocalix platform.
  • To explore fine-grained aspects of task performance derived from the timing and syntax of serial subtraction audio responses.


  • Participants were all fluent English speakers, and completed serial subtraction by three and seven, via a device-agnostic web-app on their own devices.
  • From a pool of 5,742 recordings of participants aged 17-86 years, 100 were randomly selected for manual review and scoring by trained raters using the Neurovocalix platform.
  • Time-aligned transcripts were used to derive timing measures (Box 1) and acceleration in responses were computed to capture slowing or speeding of responses as the task progresses.
  • Automated analysis of syntax from responses was used to characterise the style of responding (Box 2), and the presence and frequency of non-target responses (intrusions or self-corrections).   



  • Significant differences were found across both accuracy and timing measures depending on difficulty level, supporting the validity of this task under remote administration.
  • Difficult trials were characterised by longer, more variable, responses and longer, more variable, gaps between responses. Rate of responding increased as the task progressed.
  • No significant interaction effects for gap length, suggesting that lower rates of responding associated with more verbose styles are driven by the increased time taken to speak the answer, rather than greater thinking time between responses. 


The results of this study demonstrate the:

  • Feasibility of automated data collection for the Serial Subtraction task.
  • Derivation of novel measures of timing in assessing task performance.
  • Limitations of rate or number of responses as a performance measures given the dependence on the individual style of responding.

Future work will focus on the exploration of the changes to these timing features in the context of neurodegeneration.

Contact us 

Tags : cantab | cognitive testing | cognition | digital tools | technology | cognitive science

Author portrait

Nick Taptiklis, Head of Technology Strategy at Cambridge Cognition