Assistance State as a Condition of Functional Measurement: Distinguishing Assisted Performance from Performance at a Lower Assistance State

1. Observed Performance has More than One Source

A functional score is rarely produced by intrinsic capacity alone. As an attribution rule rather than a model: observed performance depends on the person, the assistance state, the task, the criterion, the environment, and the measurement method.

Consider two timed walks. Between them the person may have grown stronger. The person may also have changed from a cane to a walker, received cueing, changed shoes, moved from a corridor to a laboratory walkway, or been allowed a handrail. If more than one of these changed, the difference cannot be attributed to the person.

Rehabilitation measurement knows this locally. A cane or a walker alters the component times of an instrumented Timed Up and Go in healthy adults [17]. In Parkinson disease, the assisted Timed Up and Go correlated somewhat less well with self-reported mobility than the unassisted one [18]. Accelerometer step counts lose accuracy in older people using assistive devices [19]. In nonambulant stroke patients, canes, orthoses and slider shoes improved the functional mobility category while gait speed and affected-side step length were unchanged: the devices changed the function without changing the impairment [20]. The observation stays real. Its source has moved.

The same problem is less visible when the assistance is cognitive. A worker’s output with an artificial-intelligence system may be faster and better than without it. That establishes an effect on the combined system. It says nothing about the worker.

2. Assistance State is a Measurement Condition

One field has already written the rule down. The European Respiratory Society and American Thoracic Society standard for field walking tests requires that supplemental oxygen, walking aids, encouragement and track layout be recorded and, for serial comparison, held constant, because each changes the distance [21]. A six-minute walk on oxygen is not wrong. It is a different condition, and the condition travels with the number.

Assistance state is the support supplied during performance. At minimum it names the source (human, mechanical, electronic, environmental, computational); the locus, in the terms Parasuraman, Sheridan and Wickens gave automation: information acquisition, information analysis, decision selection, or action implementation [11]; the degree (full, partial, minimal, absent); the contingency (continuous, fixed, adaptive, assist-as-needed); the exposure (novel, recent, habitual); and the condition at the moment of measurement (active, reduced, absent).

One contingency needs a rule of its own. Under an adaptive or assist-as-needed controller the assistance level is a function of the person’s performance: the robot supplies less as the person supplies more [22,23]. Assistance state is then an outcome as well as a condition, and it cannot be held fixed by declaration. The remedy is the one the robotics literature already uses: report the supplied assistance, measured directly where the system permits it, as controller output or an assistance parameter, alongside the score. Without it an assist-as-needed result is uninterpretable on either axis.

Rehabilitation has graded human assistance for forty years, and every physiatrist has written the grades on an order and read them on a chart: independent, distant supervision, supervision, contact guard, minimal, moderate and maximal assist, dependent, and dependent on two. The Functional Independence Measure defined them by the person’s share of the effort, roughly three quarters, half, a quarter, and less [4]. Its successor on the American inpatient instrument, Section GG, grades the helper’s effort on a related but not identical ladder and, by design, does not score the device [32]. The Functional Ambulation Category grades walking from continuous support through light touch and supervision to independence, without pretending the shares can be counted [33]. The grades are ordinal, not measured, and the instruments do not interchange. But the grammar is settled: assistance has always been recorded as a state with a name, and the state has always traveled with the score.

Not every study needs the whole taxonomy. The condition required to interpret the result should be recorded, and the shortest useful record is the source, the degree, and the condition at measurement.

This matters most in longitudinal work. If performance improves between visits while assistance also increases, the person may have improved, the assistance may have improved, both, or the person may have declined while the assistance more than compensated. Added assistance is not itself evidence of decline; it may reflect precaution, availability, preference, or a new setting. What becomes uncertain is performance under a common assistance state. Observed performance may hold steady while the person’s share falls and the assistance supplies the difference. That is the case most likely to escape notice. Macnamara and colleagues have hypothesized that AI assistance may accelerate skill decay in experts and hinder skill development in learners while both remain unaware, because the assisted output looks unchanged [24]. Successful assistance can make a falling contribution invisible. The concern is not the assistance. It is the attribution.

3. Three Questions, Three Quantities

Three questions are usually collapsed into one.

What can the combined system do? This is assisted performance, and for many clinical purposes it is the outcome that counts.

What can the person do now when the aid is reduced or absent? This is performance at a specified lower assistance state, measured at the same time as the assisted score. Its difference from the assisted score is the orthotic effect of the assistance.

What can the person do at that lower state after a period of using the aid, compared with before? This is the question whenever assistance is meant to train, restore, or preserve rather than substitute for a capacity that is gone. It is the therapeutic effect in the language of electrical stimulation, and the retention test in the language of motor learning. I call it performance at a lower assistance state after exposure. Where the person had the ability before and still has it, the plain word is retained; where the ability is new, as it may be in an older adult whose grip has been declining, retained is the wrong word and gained is the right one. The measurement does not distinguish them; the baseline does.

The second question has two kinds of answer, and they should not be confused. Supplied assistance is what the aid delivered, and it can sometimes be measured: robot torque, body-weight unloading, the number of prompts. Required assistance is the least assistance state at which the person still meets the prespecified criterion, and it can always be observed. The person’s share of a performance need not be computed as a percentage. It is inferred to have risen when the same criterion is met with less help. Rehabilitation has operationalized assistance this way for decades: a change from moderate to minimal assistance on a fixed task is a change in assistance requirement, and it means something without any claim that the intervals are equal.

None of these is a measure of capacity. Each is a performance under a declared condition. When the task validly indexes a capacity, a performance at a lower assistance state may support an inference about that capacity, and the inference belongs to the investigator, not to the measurement.

A study of the third quantity needs an outcome at a prespecified lower assistance state after exposure, a baseline at the same state, and, where a causal claim is intended, a comparator. A post-intervention unassisted score alone shows nothing about the intervention; natural recovery, progression, practice, concurrent therapy, and selection remain.

4. Assisted, Compensated, and Restored are States, not Device Classes

The same device occupies different roles at different times. A walker supports mobility while strength and balance return. Months later its continued use may be unnecessary. Years later, after new disease, the same walker may be the correct permanent compensation. The walker did not change category. The relation among person, task, time, and assistance did.

Three states are useful. In the assisted state, performance is better or safer while help is present and the effect on later unassisted performance is unknown; most studies of assistance begin and end here. In the compensated state, assistance substitutes for a capacity that is absent or not reasonably expected to meet the task safely; compensation is not failed rehabilitation but half of it, and its goal is maximal independence with the assistance rather than withdrawal from it. In the restored state, the person performs better at a comparable lower-assistance condition after a period of use than before it.

The electrical stimulation literature keeps this distinction by population: an umbrella review found orthotic and therapeutic effects in stroke and only the orthotic effect in multiple sclerosis [9]. Same device, two states.

These states are declarations about a person performing a task at a time. They should not be assigned permanently at prescription, and where the underlying condition can change, the declaration should be renewed. Any classification needs a date.

Chuck, in his seventies, went home after a hip fracture with a walker meant for the weeks of recovery. At six weeks he walked the clinic hallway to the same chair, walker in both hands, and the note said independent with walker. At two years the hallway and the chair had not changed, and neither had the words in the chart. The walker was no longer a bridge. Nobody wrote down the visit when habit became compensation. Chuck is a composite of many patients, and the chart is every chart.

5. The Structure is Already in the Literature

The framework does not claim that investigators have never compared assisted and unassisted performance. They have, and the examples are the evidence.

Park and colleagues trained eleven chronic stroke patients with a user-driven wearable hand orthosis and assessed them with and without the device: grasping improved with assistance on, and Fugl-Meyer scores without assistance improved at the distal joints, which they present as assistive and rehabilitative effects [12]. Yurkewich and colleagues administered Goal Attainment Scaling and the Box and Block Test with and again without a hand exoskeleton after clinic and home use, and found improvement in both conditions [13]. Radder and colleagues randomized 91 older adults with declining hand function to assistive use, therapeutic use, or control of a soft robotic glove for four weeks, and reported gains in unassisted grip and pinch strength in the therapeutic group [25]; Kottink and colleagues measured unsupported grip strength after six weeks of home use of a grip-supporting glove and found it 1.9 kg higher [26]. McGibbon and colleagues, in a randomized crossover trial of the Keeogo exoskeleton in 29 people with multiple sclerosis, found small decrements in clinical performance while wearing the device and improvement in unassisted performance after two weeks of home use, 27.9 m on the six-minute walk [27]. The device-on and device-off numbers moved in opposite directions. Neither is wrong. They answer different questions.

The same separation appears with cognitive assistance. Noy and Zhang gave 453 professionals ChatGPT for writing tasks and found time down 40 percent and quality up 18 percent [28]; that is assisted performance, cleanly measured, and the experiment did not ask what the writers could do afterward without it. Bastani and colleagues ran a field experiment in roughly a thousand high-school mathematics students: an unrestricted GPT interface raised practice grades by 48 percent and a tutor configuration by 127 percent, and when access was removed the unrestricted group scored 17 percent lower on the examination than students who never had access, an effect the tutor configuration largely removed [29]. That is the guidance hypothesis, run at scale, with a language model as the guide.

Human factors reached the same result through automation. Endsley and Kiris showed that operators at higher levels of automation lost situation awareness and were slower to recover control when the automation failed [10]. Onnasch and colleagues, in a meta-analysis, found that higher degrees of automation improve routine performance and reduce workload while degrading failure performance, with the critical boundary between automation of information analysis and automation of action selection [30]. Tatasciore and colleagues showed the boundary in one task: automation that implemented the action improved performance and cut workload, and operators were less likely to notice when it failed than operators given only a recommendation [31].

These are not one phenomenon. They share one structure: performance with support, support reduced or removed, performance reassessed. That structure is the framework.

6. The Handback as a Measurement Operation

A handback is a prespecified reduction in assistance, under a fixed criterion, undertaken to determine the least assistance the person currently requires to meet it.

It is broader than withdrawal. The correct sequence is seldom full assistance to none. It is a gradient: full assistance, partial assistance, assist-as-needed, cueing only, friction removal only, and then a specified lower-assistance or unassisted performance. A rehabilitation robot reduces its share as the participant generates more of the movement; the assist-as-needed controllers built for exactly this reason, including one that deliberately decays its own assistance so the patient must re-supply the effort, are the model [22,23]. A therapist progresses from physical assistance to contact guarding to verbal cueing to independence. A cognitive system moves from answer to hint to retrieval to nothing. The question at each rung is whether the criterion is still met.

The closest existing instrument to a handback is the Walking Index for Spinal Cord Injury. It fixes the task, ten meters, and varies the combination of physical assistance, braces and devices to find the highest level the person can attain [34,35]. Kim and colleagues compared the level people selected for themselves with the maximal level they could reach. Of fifty people with chronic incomplete injury, thirty-six could walk at a less-assisted level than the one they used, twenty-one of them by three levels or more, and the habitual level was faster and cost less energy [36]. That is the distinction this paper draws, already measured in one population: performance in the ordinary assistance state, and what becomes visible when the state is changed. The handback generalizes the operation beyond walking and beyond physical help.

A handback is not universal. It is inappropriate where withdrawal creates unacceptable risk, where the compensated capacity is known to be gone, or where the assisted state is itself the clinically meaningful endpoint. A wheelchair user may be fully independent because the chair remains present.

The framework does not prescribe withdrawal. It prescribes a declaration. If an intervention is meant to restore, train, or preserve, its evaluation specifies how the person’s contribution will be assessed when assistance is safely reduced. If it is meant for permanent compensation, the important outcomes are safety, participation, autonomy, burden, satisfaction, and assisted function, and no handback is owed.

7. Criterion Lock

Changing the assistance while also changing the task produces an uninterpretable handback. Criterion lock holds the target constant while the assistance state changes: the same task, the same success criterion, the same equipment except the assistance under study, the same environment where feasible, the same measurement method, a comparable point relative to fatigue and medication, and a rule for success set in advance.

Clinical life rarely permits all of that. A patient recovering from stroke over twelve weeks may go from walker to cane, from corridor to home, learn a new gait pattern, and recover neurologically, all at once. The rule does not fail in that case; it says what can and cannot be claimed. A fully locked handback supports attribution to the person. A partially locked one supports attribution only to the set of things that moved together, and the record should say which they were. A handback with no lock at all, where the task got easier as the aid came off, supports nothing about the person; it measures the adaptation of the test.

The minimum record is therefore short. Beside every serial functional value: the task and its success criterion; the assistance state at measurement (source, degree, condition); and whether either changed since the last value. If the criterion is quietly lowered as performance declines, apparent independence is maintained by moving the mark, and the record is the only thing that will show it.

With that record, four cases cover a serial comparison. They assume the same task and criterion, and a difference judged against the test’s measurement error. Device use and human help are recorded separately; a change of device is a change of state, not a step up or down.

Assistance between visits	The comparison supports	It does not support
Unchanged	The measured performance, and its change, under the same recorded state	Unassisted performance, if assistance was used; or stability, from equal scores alone
Increased	Each measurement under its own recorded state	Attributing the change, or its absence, to the person; or inferring decline from the added help
Decreased	Performance with the support reduced, including whether the score held	Improvement under a common state; or gained capacity from reduced help alone
Not recorded at one or both visits	The two numbers and their difference	That the states were comparable; or any change in the person

Chuck’s chart sits in the first row. The recorded state did not change and neither did the result, and the row permits exactly what the chart said: independent with walker, twice. It does not permit the reading the chart invited, that nothing had changed. What changed was the walker’s role, and no row can recover a state that was never written down. Recording assistance shows when two numbers cannot be compared; it does not supply the measurement that was never taken.

8. The Help Matrix

Two axes (Figure 1). The first is the orthotic axis: at one point in time, is performance better, the same, or worse with the assistance active than at the declared lower state? The second is the therapeutic axis: within the person, is performance at the declared lower state better, the same, or worse after a period of exposure than before it?

The axes are independent, and the examples show it. Better with, better after: the wearable hand robot, whose users grasped better with it on and scored better on the Fugl-Meyer without it after training [12], and the therapeutic glove, whose users had more unassisted grip after four to six weeks of use [25,26]. Worse with, better after: the Keeogo exoskeleton, which cost performance while worn and improved unassisted walking after two weeks at home [27]; a training effect without immediate augmentation. Better with, worse after: the guidance-hypothesis cell. The direct within-person evidence is the motor-learning literature [5,7]; the AI and automation results that belong here, Bastani’s unrestricted group [29] and the higher-automation conditions in Onnasch [30], are between-group contrasts against unassisted controls, not before-and-after measurements, and the figure marks them as such. Better with, same after: augmentation or appropriate compensation with no training effect; the common case, and the one most device trials report without saying so. The remaining cells are logically required and unfilled here.

The first axis does not predict the second. That is the whole point, and it matters most when assistance is effective enough that a falling human contribution does not change the visible output.

9. Artificial Intelligence Makes the Hidden Condition Easy to See

Physical assistance is visible. A walker is in the room; a therapist’s hand can be recorded. Cognitive assistance disappears into the product. A paragraph written with a language model looks like a paragraph; a decision supported by an algorithm appears in the record as the clinician’s.

Current evaluations of AI ask whether it improves immediate output: accuracy, speed, quality, workload. Those are legitimate assisted-performance outcomes. If the longer question is human capability, another phase is required in which the system steps back far enough to ask whether the user can still do the task. Education supplies the clean case because assisted practice can be followed by an unassisted examination [29]. In professional work the equivalent is less obvious and the requirement is the same. A system can pass the first test and fail the second. That is not a reason not to use it. It is a reason to know which outcome is being optimized.

10. Implications for Trials, Longitudinal Studies, and Care

For trials: state which of the three quantities is the primary outcome; “functional improvement” is insufficient when assistance state changed. Record assistance state at every assessment. Where both assisted and lower-state outcomes matter, measure both and do not pool them. Standardize exposure before assisted testing, so that learning to operate a device is not read as the device’s effect. Include a reduction phase when retention is the question; a study of automation or AI that ends while the tool is active cannot say what the user kept. Under adaptive assistance, report the supplied assistance as the system measures it.

For care: the aid’s purpose should be explicit at prescription. What is the assistance supplying; is that capacity expected to recover, remain limited, or is its course uncertain; what defines success while the assistance is present; if recovery or retention is an objective, how and when will the person’s contribution be reassessed safely; and what result would justify maintaining, reducing, changing, or increasing the assistance. For a permanently compensatory device the fourth answer may be that no handback is required. For restorative assistance, leaving it unanswered leaves the central rehabilitation question open. A person may still cross the room. What changed may be who, or what, is doing the crossing.

11. Limitations

The framework is conceptual and proposes no score. Any difference between assisted and lower-assistance performance becomes a clinical quantity only after its reliability, practice effects, responsiveness and prognostic meaning are established for a task and population. Unassisted performance is not inherently more meaningful than assisted performance, and the framework must not be used to privilege device-free function as an ideal. A handback can be unsafe or meaningless and is used only where the question requires it. The rehabilitation grades borrowed here are ordinal. The FIM’s fractions of effort were never instrumented, WISCI levels are not equal intervals, and whether contact guarding counts as assistance is a judgment on which experienced raters differ [36]. The framework borrows the principle, not the precision. The synthesis crosses literatures with different aims; their commonality is methodological, not mechanistic. No published longitudinal series documenting the pattern described here, stable performance while the person’s share fell, was found; the clinical case is a composite, and the claim is a limit on inference, not an estimate of how often assistance conceals decline. The framework does not solve causal inference. Its purpose is narrower: to prevent a change in assistance from being mistaken for a change in the person.