Australasian Science: Australia's authority on science since 1938

The Language of Disease

By Tim Olds

What if doctors could diagnose disease by what you wrote on Facebook and Twitter?

There are so many ways doctors can diagnose disease. They can take pictures of you inside and out; they can take samples of your blood, muscles, hair, fingernails, faeces, urine or saliva; they can measure your strength, flexibility, lung function and reflexes; they can swab your skin and your intimate cavities; they can even analyse your breath and your farts.

But what if they could diagnose disease by what you wrote, by your tweets and Facebook posts, by your letters and diaries, and even the essays you wrote as a child? A whole new branch of diagnostics and epidemiology has grown up around lexical analysis — the analysis of words.

In 1986, a Minnesota psychologist started an extraordinary investigation that became known as the Nun Study. It was a longitudinal study of ageing in 678 Roman Catholic nuns. They were an interesting group because they practised celibacy and avoided smoking and alcohol.

One of the most remarkable findings of the study came from an examination of autobiographical essays the sisters wrote when they started their novitiate, at an average age of 22. By looking at the “linguistic density” of these essays – which roughly means how many ideas are packed into a small number of words – the Minnesota researchers could predict how likely the nuns were to develop Alzheimer’s disease. More than 80% of the nuns who wrote low linguistic density essays developed dementia, while only 10% of those who wrote high linguistic density essays succumbed.

The complexity of the essays probably reflected what is known as cognitive reserve – the amount of cognitive ability we build up as a child, which is probably a marker of the number of connections (synapses) between brain cells. The nuns who used more complex language had more synapses to lose, so they were more resilient to age-related decline. The more complex the language, the greater the cognitive reserve and the slower the decline.

One simple way of rating the complexity of texts is simply to count the number of words per sentence and the number of syllables per word. The most common scoring system, the Flesch-Kincaid Grade Score, estimates how many years of education a person would need to read a text easily using a simple formula:

Grade Score = 0.39 x words per sentence + 11.8 x syllables per word – 15.59

Consider the first sentence from my favourite novel, Ernst Jünger’s On the Marble Cliffs:

You all know the wild grief that besets us when we remember times of past happiness: how far beyond recall they are, and we are severed from them by something more pitiless than leagues and miles.

This has a Flesch-Kincaid Grade Level of 16, which means that you would need to have had 16 years of schooling to understand it easily. Compare this to the last text message I received:

We are watching Seb’s presentation. We have organised to meet at Sjaan’s hotel at 7. Then go to Camps Bay for dinner. We will meet at 6:50 at our hotel and we can walk over together.

This has a Flesch-Kincaid score of 4, readable by a 9-year old. This probably says something about my friends’ perceptions of my abilities.

Texts can also be analysed according to their sentiment. Lists of words are rated by groups of humans according to their valence (the balance of positive and negative emotions), arousal and dominance, and sometimes lists are extended using machine learning techniques.

One study found the most positive words to be happiness and fantastic, and the most negative to be paedophile and rapist. It’s hard to disagree with those, but less obvious words like sky and mud also have strong positive and negative valences. The most arousing words were insanity and rampage, the calmest were grain (I thought that was odd, too) and dull. The most dominant words were incredible and self, the least dominant dementia and lobotomy.

An overall rating is calculated for a text such as a website, tweet, political speech, popular song or written piece. In the Nun Study, novices whose essays were rated in the bottom 25% for positive emotions had three times the risk of dying at any given age as those whose essays were rated in the top 25%.

Mass lexical analysis has been widely used in business to assess the popularity and brand image of various products, but it can also be applied to health. One study analysed Google searches for mentions of flu symptoms and medications from 2003–08 in the US. Using geolocation, the authors managed to reproduce almost exactly the incidence of flu as determined by the Centers for Disease Control – and they did it 1–2 weeks before the CDC. Another study mined the Twitter corpus to find links between diseases and everyday life, finding a strong link between the word pimples and depression. Another UK study has mined tweets to explore adverse drug reactions.

The world is awash with words. They leave an enduring trace, and lexical analysis can chart the history and predict the future of our health. And in case you’re wondering, the Flesch-Kincaid Grade Score for this article is 14. You see, I have a high opinion of the linguistic capabilities of readers of Australasian Science.

Professor Tim Olds leads the Health and Use of Time Group at the Sansom Institute for Health Research, University of South Australia.